TWI823256B - Light color coordinate estimation system and deep learning method thereof - Google Patents

Abstract

A light color coordinate estimation system and a deep learning method thereof are disclosed. The light color coordinate estimation system includes a plurality of light detectors, each of which has different spectral responses, a normalized calculation circuit and a neural network. The light detectors respectively detect integrated energy intensities of a light beam which color is to be measured. The normalized calculation circuit is electrically connected to the light detectors and respectively divides the maximum value obtained by light detectors. The neural network is electrically connected to the normalized calculation circuit to obtain the normalized values and output an estimated color coordinate the light beam to be measured.

Description

Light source color coordinate estimation system and its deep learning method

The present invention relates to a light source color coordinate estimation system and a deep learning method thereof, and in particular to a light source color coordinate estimation system based on a neural network architecture and a deep learning method thereof.

Different color coordinates represent different color temperatures and their color characteristics respectively. The current method of obtaining the color coordinates of a light source is to obtain the RGB values of the light source through a sensor, and then multiply the RGB values by a conversion matrix to calculate the color coordinate values.

However, the conversion of RGB values into color coordinates can produce errors. In addition, some light sources have brighter surfaces with higher reflectivity, while some light sources have rougher surfaces with lower reflectivity. Differences in reflectivity of the light source surface and differences in the intensity of the beam emitted by the light source can also lead to errors in the calculation of color coordinates.

The technical problem to be solved by the present invention is to provide a deep learning-based light source color coordinate estimation system and its deep learning method in view of the shortcomings of the existing technology.

In order to solve the above technical problems, one of the technical solutions adopted by the present invention is to provide a light source color coordinate estimation system, which includes a plurality of light detectors, a normalized calculation circuit, and a neural network. The light detectors have respective spectral responses after receiving the light beam emitted by the light source, where the spectral response includes one or more detection bands and multiple energy integrated values corresponding to the detection bands. The normalized calculation circuit is electrically connected to the light detectors, and divides the energy integrated values by the maximum of the energy integrated values to calculate a plurality of normalized energy integrated values. The input end of the neural network is electrically connected to the normalized computing circuit. The neural network includes multiple neurons, which are connected through multiple synapses. Each synapse has a weight value, at least partially The neurons include an excitation function. The input end of the neural network receives the normalized energy integral values and the normalized energy integral values are converted into estimated color coordinates through the operation of the excitation functions and the weight values. The neural network The output terminal of the path outputs the estimated color coordinates.

In order to solve the above technical problems, another technical solution adopted by the present invention is to provide a deep learning method for a light source color coordinate estimation system, which includes: providing a plurality of different light sources, and executing a deep learning program on these light sources in sequence. , the deep learning program includes: obtaining multiple energy integral values of multiple different detection bands in the light source through multiple light detectors; dividing all energy integral values by all energy integral values through a normalized calculation circuit to calculate multiple normalized energy integral values; receive these normalized energy integral values through the neural network, where the neural network includes multiple excitation functions and multiple weight values; according to the excitations of the neural network The calculation of the function and the weight values converts the normalized energy integral values into estimated color coordinates; calculates the color coordinate error between the estimated color coordinates and the preset color coordinates of the light source; and based on the back-propagation calculation of the neural network method and coordinate error to adjust at least one weight value of the neural network.

In one embodiment, the sum of the detection bands of all light detectors preferably covers the entire visible light region (380nm~700nm), the input layer includes a plurality of input layer neurons, and the number of the input layer neurons Same as the number of light detectors, the input layer neurons obtain the normalized energy integral values respectively.

One of the beneficial effects of the present invention is that the light source color coordinate estimation system and its deep learning method provided by the present invention can obtain the normal results of different light detectors because the neural network performs deep learning on the induction differences of different light detectors. The integrated energy value, in which the sum of the spectral responses of all light detectors preferably covers the entire visible light region (380nm~700nm). Therefore, the neural network has a high anti-interference ability, making the estimated color coordinates highly accurate. Furthermore, the normalization calculation circuit normalizes the values of multiple light detectors every time the light source emits a beam. In this way, the integrated energy value detected by each light detector will not be weakened by the intensity of the light beam emitted by the light source or the reflectivity of the surface of the light source, which also makes the estimated color coordinates more accurate. Spend.

In order to further understand the features and technical content of the present invention, please refer to the following detailed description and drawings of the present invention. However, the drawings provided are only for reference and illustration and are not used to limit the present invention.

The following is a specific embodiment to illustrate the implementation of the "light source color coordinate estimation system and its deep learning method" disclosed in the present invention. Those skilled in the art can understand the advantages and effects of the present invention from the content disclosed in this specification. The present invention can be implemented or applied through other different specific embodiments, and various details in this specification can also be modified and changed based on different viewpoints and applications without departing from the concept of the present invention. In addition, the drawings of the present invention are only simple schematic illustrations and are not depictions based on actual dimensions, as is stated in advance. The following embodiments will further describe the relevant technical content of the present invention in detail, but the disclosed content is not intended to limit the scope of the present invention.

It should be understood that although terms such as “first”, “second” and “third” may be used herein to describe various elements or signals, these elements or signals should not be limited by these terms. These terms are primarily used to distinguish one component from another component or one signal from another signal. In addition, the term "or" used in this article shall include any one or combination of more of the associated listed items depending on the actual situation.

[First Embodiment]

FIG. 1 is a functional block diagram of a light source color coordinate estimation system according to the first embodiment of the present invention. Referring to Figure 1, the light source color coordinate estimation system 100 includes a plurality of light detectors 1A-1F, a normalization calculation circuit 2 and a neural network 3. The light detectors 1A-1F respectively have one or more A detection waveband is used to detect energy integral values corresponding to different detection wavebands in a light beam emitted by a light source S, wherein the sum of these detection wavebands preferably covers the entire visible light region (380nm~700nm). The light detectors 1A to 1F are electrically connected to the normalization calculation circuit 2 respectively, and the normalization calculation circuit 2 is further electrically connected to the neural network 3. The neural network 3 outputs the first estimated color coordinate value x and The second estimated color coordinate is y, and the first estimated color coordinate value x and the second estimated color coordinate value form an estimated color coordinate (x, y). Through the estimated color coordinate (x, y), the color of the light emitted by the light source can be known.

FIG. 2 is a spectral response diagram of the photodetectors 1A to 1F in FIG. 1 . As shown in Figure 2, the unit of the horizontal axis is wavelength (nm), and the unit of the vertical axis is light energy. The curves SPR1 to SPR6 respectively represent the six spectral responses of the light detectors 1A-1F. Each spectral response reflects It is the detection band and energy intensity of the light detector under beam irradiation. The detection wavebands of these photodetectors are all different, that is, the detection wavebands of any two photodetectors can partially overlap but not be completely consistent, so as to increase the diversity of the sensing range as much as possible. Each of the photodetectors 1A-1F sums up the light energy in the corresponding detection band to obtain a corresponding energy integral value. As shown in Figure 2, the sum of these detection bands is exemplified as 380nm to 730nm, covering the visible light region (380nm~700nm) to avoid inaccurate detection results due to depletion areas in the visible light band.

Regarding the implementation of the photodetectors 1A-1F, for example, it is implemented using a hardware device. The hardware device includes a circuit board, a diode, a metal semi-field effect transistor (MOS), and a bipolar junction resistor. Any combination of crystal (BJT), resistor, inductor, capacitor. Alternatively, it can also be implemented through firmware, editing a hardware language (such as VHDL) and burning the hardware language into a microcontroller (MCU) or field programmable gate array (FPGA). Alternatively, it can be implemented through software, such as editing C language.

Referring again to Figure 1, the normalized calculation circuit 2 includes six input terminals 21A~21F, a central processing circuit 23, a non-volatile memory 25 (Non-Volatile Memory) and six output terminals 27A~27F. The six input terminals are The terminals 21A to 21F are electrically connected to the photodetectors 1A to 1F respectively, and the central processing circuit 23 is electrically connected to the six input terminals 21A to 21F, the non-volatile memory 25 and the six output terminals 27A to 27F. The normalized calculation circuit 2 obtains six different energy integral values through the six input terminals 21A to 21F. The non-volatile memory 25 is a read-only memory (ROM), a flash memory (Flash memory) or a non-volatile random access memory (RVRAM) and stores a normalization algorithm. Regarding the normalization algorithm, the maximum value is found from N different values, and then the N values are divided by the maximum value to calculate N normalized values.

In this embodiment, the central processing circuit 23 reads and executes the normalization algorithm stored in the non-volatile memory 25 in order to normalize the six energy integral values, which is found among the six energy integral values. The maximum value is then divided by the maximum value to calculate the six normalized energy integral values.

As for the implementation of the normalized computing circuit 2, for example, it is implemented using hardware devices. The hardware devices include circuit boards, diodes, metal semi-field effect transistors (MOS), and bicarrier junction transistors ( BJT), resistors, inductors, capacitors in any combination. Alternatively, it can also be implemented through firmware, editing a hardware language (such as VHDL) and burning the hardware language into a microcontroller (MCU) or field programmable gate array (FPGA). Alternatively, it can be implemented through software, such as editing C language.

The normalization processing of the normalization calculation circuit 2 produces at least the following beneficial effects: the normalization calculation circuit 2 normalizes the integrated energy values detected by multiple light detectors every time the light source S emits a beam. In this way, the energy integrated value of the light detector will not be weakened due to differences in the intensity of the light beam emitted by the light source or differences in reflectivity of the light source surface.

Referring again to Figure 1, the neural network 3 includes an input layer 31, a first hidden layer 33, a second hidden layer 35 and an output layer 37. In this embodiment, the input layer 31 includes six input layer neurons. The six output terminals 27A to 27F of the calculation circuit 2 are respectively connected to the six input layer neurons of the input layer 31, so that the six input layer neurons of the input layer 31 obtain six normalized energy integral values respectively.

In this embodiment, the first hidden layer 33 includes eight hidden layer neurons. Each input layer neuron of the input layer 31 is connected to the eight hidden layer neurons of the first hidden layer 33 via eight synapses. Therefore, The input layer 31 and the first hidden layer 33 are connected via a total of forty-eight synapses, and each of the forty-eight synapses has a specific weight value. Therefore, the output value of each input layer neuron of the input layer 31 is multiplied by eight weight values and then transmitted to the eight hidden layer neurons of the first hidden layer 33 respectively. Each hidden layer neuron of the first hidden layer 33 uses a preset activity function to operate on the input value to generate an output value, and the activity function (activity function) is, for example, a sigmoid function or a relu function, but is not based on This is the limit.

Regarding the calculation between the input layer 31 and the first hidden layer 33, for example, the six input layer neurons of the input layer 31 are connected to the first hidden layer neuron of the first hidden layer 33 via six synapses. connection, and the weight values of these six synapses are w1~w6 respectively, and the output values of the six input layer neurons of the input layer 31 are x1~x6 respectively. At this time, the first hidden layer neuron of the first hidden layer 33 The input value y1 received by the input terminal of the element is = x1*w1+x2*w2+x3*w3+x4*w4+x5*w5+x6*w6, which is used by the first hidden layer neuron of the first hidden layer 33. The activation function is ,and . Therefore, the output value of the first hidden layer neuron of the first hidden layer 33 . According to the above example calculation formula, the output values of the other seven hidden layer neurons of the first hidden layer 33 can be deduced in the same way.

The second hidden layer 35 in this embodiment includes four hidden layer neurons, and each hidden layer neuron of the first hidden layer 33 is connected to the four hidden layer neurons of the second hidden layer 35 via four synapses. Therefore, the first hidden layer 33 and the second hidden layer 35 are connected through a total of thirty-two synapses, and each of the thirty-two synapses has a specific weight value. Therefore, the output value of each hidden layer neuron of the first hidden layer 33 is first multiplied by four weight values and then transmitted to the four hidden layer neurons of the second hidden layer 35 respectively. Each hidden layer neuron of the second hidden layer 35 uses a preset activity function to operate on the input value to generate an output value.

Regarding the calculation between the first hidden layer 33 and the second hidden layer 35, for example, the eight hidden layer neurons of the first hidden layer 33 are connected to the first hidden layer of the second hidden layer 35 via eight synapses. layers of neurons are connected, and the weight values of these eight synapses are w7~w14 respectively. The output values of the eight hidden layer neurons of the first hidden layer 33 are x7~x14 respectively. At this time, the output values of the eight hidden layer neurons of the second hidden layer 35 are The input value y2 received by the input terminal of the first hidden layer neuron=x7*w7+x8*w8+x9*w9+x10*w10+x11*w11+x12*w12+x13*w13+x14*w14, No. The activation function used by the first hidden layer neuron of the second hidden layer 35 is ,and . Therefore, the output value of the first hidden layer neuron of the second hidden layer 35 is . According to the above example calculation formula, the output values generated by the other three hidden layer neurons of the second hidden layer 35 can be deduced in the same way.

In this embodiment, the output layer 37 includes two output layer neurons. Each hidden layer neuron of the second hidden layer 35 is connected to the two output layer neurons of the output layer 37 via two synapses. Therefore, the second hidden layer 35 and the output layer 37 are connected via a total of eight synapses, and each of the eight synapses has a specific weight value. Therefore, the output value of each hidden layer neuron of the second hidden layer 35 is first multiplied by two weight values and then transmitted to the two output layer neurons of the output layer 37 respectively. Each output layer neuron of the output layer 37 uses an activity function to operate on the input value to generate an output value. The two output layer neurons of the output layer 37 respectively output the first estimated color coordinate value x and the second estimated color coordinate value y.

Regarding the calculation between the second hidden layer 35 and the output layer 37, for example, the four hidden layer neurons of the second hidden layer 35 are connected to the first output layer neuron of the output layer 37 via four synapses. , and the weight values of the four neural chains are w15~w18 respectively, and the output values of the four hidden layer neurons of the second hidden layer 35 are x15~x18 respectively. At this time, the input of the first output layer neuron of the output layer 37 The input value y3= x15*w15+x16*w16+x17*w17+x18*w18 received by the terminal, the excitation function used by the first output layer neuron of output layer 37 is ,and . Therefore, the output value of the first output layer neuron of output layer 37 is . According to the above example calculation formula, the output value of another output layer neuron of the output layer 37 can be deduced in the same way.

As for the implementation of neural network 3, for example, it is implemented using hardware devices. The hardware devices include circuit boards, diodes, metal semi-field effect transistors (MOS), and bipolar junction transistors (BJT). ), any combination of resistors, inductors, and capacitors. Alternatively, it can also be implemented through firmware, editing a hardware language (such as VHDL) and burning the hardware language into a microcontroller (MCU) or field programmable gate array (FPGA). Alternatively, it can be implemented through software, such as editing C language.

[Second Embodiment]

FIG. 3 is a functional block diagram of a light source color coordinate estimation system according to the second embodiment of the present invention. The difference between the light source color coordinate estimation system 200 of the second embodiment of FIG. 3 and the light source color coordinate estimation system 100 of the first embodiment of FIG. 1 is that the light source color coordinate estimation system 200 includes eight light detectors 1A -1H, the normalization calculation circuit 2 includes eight input terminals 21A~21H, the eight input terminals 21A~21H are electrically connected to the photodetectors 1A~1H respectively, and the central processing circuit 23 is electrically connected to the eight inputs Terminals 21A～21H. The central processing circuit 23 reads and executes the normalization algorithm stored in the non-volatile memory 25 in order to normalize the eight energy integrated values detected by the eight light detectors. It first starts from the eight Find the maximum value among the energy integral values, and then divide the eight energy integral values by the maximum value to calculate eight normalized energy integral values. The input layer 31 of the neural network 3 includes eight neurons, and the eight output terminals 27A~27H of the normalized calculation circuit 2 are respectively connected to the eight neurons of the input layer 31, so that the eight input layers of the input layer 31 Each neuron obtains eight normalized energy integral values.

[Third Embodiment]

FIG. 4 is a functional block diagram of a light source color coordinate estimation system according to the third embodiment of the present invention. The difference between the light source color coordinate estimation system 300 of the third embodiment of FIG. 4 and the light source color coordinate estimation system 100 of the first embodiment of FIG. 1 is that the hidden layer of the first hidden layer 33 of the light source color coordinate estimation system 300 The number of neurons is changed to five and the number of hidden layer neurons of the second hidden layer 35 is changed to three. Each hidden layer neuron of the first hidden layer 33 is connected to the six hidden layer neurons of the input layer 31 via six synapses. There are a total of three synapses connected between the first hidden layer 33 and the input layer 31 . Ten, and each synapse has a specific weight value. Each hidden layer neuron of the second hidden layer 35 is connected to the five hidden layer neurons of the first hidden layer 33 via five synapses, and is connected to the synapses between the first hidden layer 33 and the second hidden layer 35 There are a total of fifteen synapses, and each synapse has a specific weight value.

[Fourth Embodiment]

FIG. 5 is a functional block diagram of a light source color coordinate estimation system according to the fourth embodiment of the present invention. The difference between the light source color coordinate estimation system 400 of the fourth embodiment of FIG. 5 and the light source color coordinate estimation system 100 of the first embodiment of FIG. 1 is that the neural network 3 includes an input layer 31 and a first hidden layer 33 , the second hidden layer 35 and the output layer 37, a third hidden layer 39 is also included. The third hidden layer 39 is between the second hidden layer 35 and the output layer 37. The third hidden layer 39 includes four hidden layer neurons. Each hidden layer neuron of the second hidden layer 35 is connected to the output layer via four synapses. The four hidden layer neurons of the third hidden layer 39 are connected. There are a total of sixteen synapses connected between the second hidden layer 35 and the third hidden layer 39, and each synapse has a specific weight value. . Each hidden layer neuron of the third hidden layer 39 is connected to two output layer neurons of the output layer 37 via two synapses. The number of synapses connecting the third hidden layer 39 and the output layer 37 is eight in total. bars, and each synapse has a specific weight value.

It can be seen from the embodiments of the light source color coordinate estimation system in Figures 1, 3, 4 and 5 that the number of detectors, the number of neurons in the hidden layer, the number of hidden layers and the excitation functions used by the neurons can be determined according to Moderate adjustments can be made based on usage requirements and the accuracy of the estimated color coordinates, and are not limited to the above embodiment.

Figure 6 shows the deep learning method of the light source color coordinate estimation system according to the first embodiment of the present invention. As shown in Figure 6, in step S601, multiple different light sources are provided, and deep learning procedures are executed on these light sources in sequence. For example, when there are ten light sources, each light source emits a light beam to the light source color coordinate estimation system, so that the light source color coordinate estimation system obtains a piece of training data. When the number of all light sources is ten, the light source color coordinate system obtains a total of ten training data and performs a total of ten deep learning. As for the deep learning program, it contains at least the following steps:

In step S603, multiple light detectors each having different detection wavebands are used to detect multiple energy integral values corresponding to the detection wavebands in the light beam emitted by the light source, and the sum of the detection wavebands is Best to cover the entire visible light region (380nm~700nm). For example, there are five light detectors, and the light detectors respectively correspond to the detection wavebands of red light, green light, blue light, yellow light and purple light. These detection wavebands partially overlap and are in The sum can cover the entire visible light region (380nm~700nm).

In step S605, the normalized calculation circuit 2 divides each energy integrated value by the maximum of all energy integrated values to calculate a plurality of normalized energy integrated values. For example, in a light detector that detects red light, green light, blue light, yellow light, and purple light, taking the energy integral value corresponding to purple light as the maximum value, the normalized calculation circuit will correspond to red light, The energy integral values of the light detectors of green light, blue light, yellow light and purple light are respectively divided by the energy integral value of purple light to calculate the Wubi normalized energy integral values respectively.

In step S607, all normalized energy integral values are received through the neural network 3, where the neural network 3 includes multiple excitation functions and multiple weight values.

In step S609, the normalized energy integral values are converted into estimated color coordinates according to the calculation of the activation functions and the weight values of the neural network 3;

In step S611, the color coordinate error between the estimated color coordinates and the preset color coordinates of the light source is calculated. For example, the estimated color coordinate x and estimated color coordinate y calculated by the light source color coordinate estimation system are 0.3 and 0.2 respectively, and the preset color coordinate x and preset color coordinate y of the light source are 0.25 and 0.15 respectively. The x-axis color coordinate error is -0.5, while the y-axis color coordinate error is -0.05.

In step S613, the weight value of the neural network 3 is adjusted according to the back propagation algorithm of the neural network 3 and the color coordinate error. For example, the input layer, hidden layer and output layer of a neural network have i neurons, j neurons and k neurons respectively. is the estimated value of the output layer, is the target value, E is the error function, where E= . The backpropagation algorithm adjusts the weight values of each neural chain connected to the input layer and hidden layer. and the weight values of each neural chain connected between the hidden layer and the output layer. To minimize the error function E, the weight adjustment formula of each neural chain connected between the input layer and the hidden layer is: , the weight value adjustment formula of each neural chain connected between the hidden layer and the output layer is: ,in The number of neurons in the input layer, the number of neurons in the hidden layer, the number of hidden layers and the learning rate The settings will affect the training effect of the neural network.

Figure 7 shows the deep learning method of the light source color coordinate estimation system according to the second embodiment of the present invention. The difference between the deep learning method of Figure 7 and the deep learning method of Figure 6 is that the deep learning method of Figure 7 further includes step S715. As for Steps S701 to S713 in Figure 7 are the same as steps S601 to S613 in Figure 6 . Regarding step S715, after each light source completes the deep learning process, it determines whether the color coordinate errors are within the preset error convergence interval. When one of the color coordinate errors is not within the error convergence interval, multiple different color coordinate errors are provided. light source to execute deep learning programs. For example, twenty different light sources were originally prepared for deep learning training of the light source color coordinate estimation system, and the preset error convergence interval is -2%~2%. When the percentage of the color coordinate error of a light source is within the error convergence range When outside the interval, other light sources different from the original twenty light sources are prepared to perform deep learning training on the light source color coordinate estimation system.

When the color coordinate errors are within the error convergence interval, the deep learning process is completed. After completing the deep learning program, a test light source can be provided and the test light source emits a test beam. The trained light source color coordinate estimation system obtains the test beam and generates the estimated color coordinate of the test beam, and the calculated estimated color coordinate of the test beam The error with the actual recorded color coordinates of the test light source is within -2%~2%, which meets the requirements for use.

8A to 8B respectively show the x-axis color coordinate error relationship diagram and the y-axis color coordinate error relationship diagram generated by the light source color coordinate estimation system for multiple different light sources. As shown in Figure 8A~Figure 8B, there are a total of twenty different light sources. The twenty light sources sequentially emit a beam of light to the light source color coordinate estimation system. The beam emitted by each light source has a preset x-axis color coordinate. As well as the y-axis color coordinate, the twenty x-axis estimated color coordinates calculated by the light source color coordinate estimation system will be compared with the preset twenty x-axis color coordinates to generate twenty x-axis colors as shown in Figure 8A The coordinate error percentage is distributed between -5% and 5%, which is in line with the standards of actual use. In the same way, the twenty y-axis estimated color coordinates calculated by the light source color coordinate estimation system will be compared with the preset twenty y-axis color coordinates to generate the twenty y-axis color coordinate error percentages as shown in Figure 8B , which is distributed between -5% and 5%, in line with the standards of actual use.

Figure 9 is a schematic diagram of the first usage state of the light source color coordinate estimation system of the present invention. As shown in FIG. 9 , the light detectors 1A to 1F are provided on the user device T1 , and the user device T1 is, for example, the user's mobile device, notebook computer, wearable device, etc. The normalized computing circuit 2 and the neural network 3 are installed on the remote host T2, and the remote host T2 is, for example, a server on the server side or a cloud computer. The user device T1 is communicatively connected to the remote host T2. Such a configuration allocates the more complex calculations in the light source color coordinate estimation system to the hardware of the remote host T2 for processing.

Figure 10 is a schematic diagram of the second usage state of the light source color coordinate estimation system of the present invention. As shown in Figure 10, the light detectors 1A~1F, the normalized calculation circuit 2 and the neural network 3 are provided in the system chip T3, and the system chip T3 can be provided in the user's mobile device, notebook computer, wearable Devices etc.

[Beneficial effects of the embodiment]

One of the beneficial effects of the present invention is that in the light source color coordinate estimation system and its deep learning method provided by the present invention, since the neural network performs deep learning on the induction differences of different light detectors, these light detectors respectively have Different detection wavebands, and the sum of these detection wavebands preferably covers the entire visible light region (380nm~700nm). Therefore, the neural network has a high anti-interference ability, making the estimated color coordinates highly accurate. Furthermore, the normalization calculation circuit normalizes multiple energy integrated values every time the light source emits a beam. In this way, the energy integrated value of each light detector will not be weakened by the intensity of the light beam emitted by the light source or the reflectivity of the surface of the light source, which also makes the estimated color coordinates more accurate. . Furthermore, since the neural network directly estimates color coordinates, errors caused by converting RGB values into color coordinates are avoided.

The contents disclosed above are only preferred and feasible embodiments of the present invention, and do not limit the scope of the patent application of the present invention. Therefore, all equivalent technical changes made by using the description and drawings of the present invention are included in the application of the present invention. within the scope of the patent.

100, 200, 300, 400: Light source color coordinate estimation system 1A~1H: Light detector 2: Normalized calculation circuit 21A~21H: Input terminal 23:Central processing circuit 25:Non-volatile memory 27A~27H: Output terminal 3: Neural Network 31:Input layer 33: First hidden layer 35:Second hidden layer 37:Output layer 39: The third hidden layer x: first estimated color coordinate value y: second estimated color coordinate value S: light source T1: User device T2: remote host T3: system chip S601~S613: steps S701~S715: steps

FIG. 1 is a functional block diagram of a light source color coordinate estimation system according to the first embodiment of the present invention.

FIG. 2 is a plot of spectral response versus wavelength for the six photodetectors of FIG. 1 .

FIG. 3 is a functional block diagram of a light source color coordinate estimation system according to the second embodiment of the present invention.

FIG. 4 is a functional block diagram of a light source color coordinate estimation system according to the third embodiment of the present invention.

FIG. 5 is a functional block diagram of a light source color coordinate estimation system according to the fourth embodiment of the present invention.

Figure 6 shows the deep learning method of the light source color coordinate estimation system according to the first embodiment of the present invention.

Figure 7 shows the deep learning method of the light source color coordinate estimation system according to the second embodiment of the present invention.

Figures 8A and 8B respectively show the x-axis color coordinate error distribution chart and the y-axis color coordinate error distribution chart of the light source color coordinate estimation system.

Figure 9 is a schematic diagram of the first usage state of the light source color coordinate estimation system of the present invention.

Figure 10 is a schematic diagram of the second usage state of the light source color coordinate estimation system of the present invention.

100:Light source color coordinate estimation system

1A~1F: Light detector

2: Normalized calculation circuit

21A~21F: Input terminal

23:Central processing circuit

25:Non-volatile memory

27A~27F: Output terminal

3: Neural Network

31:Input layer

33: First hidden layer

35:Second hidden layer

37:Output layer

x: first estimated color coordinate value

y: second estimated color coordinate value

S: light source

Claims (10)

A light source color coordinate estimation system includes: a plurality of light detectors. The light detectors respectively have different detection wavelength bands. When the light detectors receive a light beam emitted by a light source, each light detector A spectral response of each detection device, and the spectral responses include the detection bands and a plurality of energy integrated values corresponding to the detection bands; a normalized calculation circuit electrically connected to the light detectors , the normalized calculation circuit divides the energy integral values by the maximum of the energy integral values to calculate a plurality of normalized energy integral values; and a neural network, an input end of the neural network Electrically connected to the normalized computing circuit, the neural network includes a plurality of neurons, and the neurons are connected through a plurality of synapses. Each synapse has a weight value, and at least some of the neurons include An excitation function, the input end of the neural network receives the normalized energy integral values and the normalized energy integral values are converted into an estimated color coordinate through the operation of the excitation functions and the weight values, the neural network An output end of the path outputs the estimated color coordinate; the neurons include multiple input layer neurons, the input layer neurons are located in an input layer of the neural network, and the number of the input layer neurons is the same as the According to the number of light detectors, the input layer neurons obtain the normalized energy integral values respectively. The light source color coordinate estimation system as described in claim 1, wherein the sum of the detection bands covers 380nm~700nm. The light source color coordinate estimation system as described in claim 1, wherein the neural network includes an input layer, a plurality of hidden layers and an output layer, and the input layer and the output layer The outgoing layer is respectively connected to two of the hidden layers, and any two adjacent ones of the hidden layers are connected to each other. The neurons include multiple hidden layer neurons, and each hidden layer includes the hidden layer neurons. part of it. The light source color coordinate estimation system as described in claim 3, wherein the activation function used by the hidden layer neuron is a sigmoid function. The light source color coordinate estimation system of claim 1, wherein the light detectors are provided in a user device, and the normalization calculation circuit and the neural network are provided in a remote host. The light source color coordinate estimation system as claimed in claim 1, wherein the light detectors, the normalization calculation circuit and the neural network are provided in a system chip. A deep learning method for a light source color coordinate estimation system, including: sequentially executing a deep learning program on multiple different light sources, wherein the deep learning program includes: receiving light beams emitted by the light source through multiple light detectors, wherein the These light detectors respectively have multiple different detection bands; when the light detectors receive the light beam, each light detection device has its own spectral response, and these spectral responses include the detection bands and Corresponding to a plurality of energy integral values of the detection bands; dividing the energy integral values by the largest of the energy integral values through a normalization calculation circuit to calculate a plurality of normalized energy integral values; Receive the normalized energy integral values through a neural network, where the neural network includes multiple neurons, multiple activation functions and multiple weight values; Convert the normalized energy integral values into an estimated color coordinate according to the calculation of the activation functions and the weight values of the neural network; calculate the difference between the estimated color coordinate and a preset color coordinate of the light source a color coordinate error; and adjusting at least one weight value of the neural network according to a backpropagation algorithm of the neural network and the coordinate error; the neurons include multiple input layer neurons, and the input layer neurons The element is located in an input layer of the neural network, the number of the input layer neurons is the same as the number of the light detectors, and the input layer neurons respectively obtain the normalized energy integral values. The deep learning method of the light source color coordinate estimation system as described in claim 7, wherein the sum of the detection bands covers 380nm~700nm. The deep learning method of the light source color coordinate estimation system as described in claim 7 further includes determining whether the color coordinate errors are in an error convergence interval after each light source completes the deep learning program. When the color coordinate errors If one of the light sources exceeds the error convergence interval, the deep learning program is executed for multiple different light sources. The deep learning method of the light source color coordinate estimation system as described in claim 9, wherein when the color coordinate errors are within the error convergence interval, the deep learning program is completed; after completing the deep learning program, a The light source is tested and the test light source emits a test beam. The light source color coordinate estimation system obtains the test beam and generates an estimated color coordinate of the test beam.

Images