JP2019161665A - Feature extraction element, feature extraction system, and determination apparatus - Google Patents

Abstract

A throughput of an entire feature extraction apparatus is improved.
A feature extraction element comprising: a light receiving substrate in which a plurality of light receiving elements for photoelectrically converting received light are two-dimensionally arranged; and one or more other substrates stacked on the light receiving substrate, The other substrate has a plurality of multiplication circuits provided corresponding to each light receiving element or each block including a plurality of light receiving elements, and convolves signals output from the plurality of light receiving elements using the plurality of multiplication circuits. A convolution processing unit for calculating, a pooling processing unit for sampling a signal output from the convolution processing unit based on a predetermined condition, and a connection wiring for passing the sampled signal to a plurality of multiplication circuits. Have.
[Selection] Figure 1

Description

  The present invention relates to a feature extraction element, a feature extraction system, and a determination device.

There is a method of speeding up processing by dividing an image into blocks and extracting features for each block (see, for example, Patent Document 1).
Patent Document 1 Japanese Patent Application Laid-Open No. 2008-148298

  Since data extracted as an image is processed and feature extraction is performed, it takes time to generate an image, transfer the generated image, and the like, and it cannot be said that the speed is sufficiently increased.

  The first aspect of the present invention includes a light receiving substrate in which a plurality of light receiving elements that photoelectrically convert received light are two-dimensionally arranged, and one or more other substrates stacked on the light receiving substrate. The substrate has a plurality of multiplication circuits provided corresponding to each light receiving element or each block composed of a plurality of light receiving elements, and convolves the signals output from the plurality of light receiving elements using the plurality of multiplication circuits. A convolution processing unit, a pooling processing unit that samples a signal output from the convolution processing unit based on a predetermined condition, and a connection wiring for passing the sampled signal to a plurality of multiplication circuits An extraction element is provided.

  In a second aspect of the present invention, the feature extraction element, a convolution operation of the convolution processing unit, and a control unit that controls to repeat sampling of the pooling processing unit are provided, and the control unit repeatedly performs the convolution operation. In this case, a feature extraction system is provided that controls the convolution processing unit to use predetermined filter coefficients.

  In a third aspect of the present invention, there is provided a determination device including the feature extraction element and a determination unit that determines an imaging target based on a feature amount extracted based on an output from a pooling processing unit.

  The above summary of the present invention does not enumerate all of the features of the present invention. A sub-combination of these feature groups can also be an invention.

2 is a schematic cross-sectional view of the feature extraction device 100. FIG. 5 is a flowchart of processing executed by the feature extraction device 100. 2 is a schematic cross-sectional view of the feature extraction device 100. FIG. 2 is a schematic cross-sectional view of the feature extraction device 100. FIG. 3 is a timing chart of the feature extraction device 100. 2 is a block diagram of an imaging apparatus 500. FIG. It is a typical sectional view of other feature extraction devices 101. 2 is a schematic cross-sectional view of a feature extraction device 101. FIG. 3 is a partial timing chart of the feature extraction apparatus 101.

  Hereinafter, the present invention will be described through embodiments of the invention, but the following embodiments do not limit the invention according to the claims. In addition, not all the combinations of features described in the embodiments are essential for the solving means of the invention.

  FIG. 1 is a schematic cross-sectional view of the feature extraction apparatus 100. The feature extraction apparatus 100 includes a microlens 110, a pixel substrate 120, an AD conversion circuit board 130, a multiplication circuit board 140, an addition circuit board 150, a convolution calculation result addition board 160, an activation function calculation circuit board 170, pooling, which are sequentially stacked. This is a single element including the circuit board 180 and the coupling circuit board 190.

  Note that the feature extraction apparatus 100 extracts a feature amount used when determining a subject from the received subject luminous flux. Examples of the determination of the subject include, but are not limited to, determining what the subject is, determining the category of the subject, and determining the degree of the subject (for example, evening scene degree). The feature amount extracted here corresponds to a feature extracted by a multilayer neural network called deep learning or deep learning. In other words, the feature extraction apparatus 100 can also be used for learning processing for extracting features by deep learning.

  Each of the substrates in the feature extraction apparatus 100 includes a wiring layer 12 formed on the base substrate 11 by photolithography. Each wiring layer 12 includes a circuit formed of wiring, elements, and the like. That is, the pixel substrate 120, the AD conversion circuit board 130, the multiplication circuit board 140, the addition circuit board 150, the convolution calculation result addition board 160, the activation function calculation circuit board 170, the pooling circuit board 180, and the coupling circuit board 190 are respectively The pixel 121, the AD conversion circuit 131, the multiplication circuit 141, the addition circuit 151, the addition circuit 161 with the latch B, the activation function calculation circuit 171, the pooling circuit 181, and the coupling circuit 191 are included.

  In addition, the substrate is electrically connected to another substrate stacked adjacently through a through electrode 13 penetrating each base substrate 11. Further, some of the substrates are directly electrically connected to non-adjacent substrates by through electrodes 14 formed through the plurality of substrates. This point will be described with reference to FIGS.

  In the feature extraction device 100, the pixel substrate 120 has a plurality of pixels 121 arranged two-dimensionally and periodically. Each of the pixels 121 includes a light receiving element such as a photodiode that photoelectrically converts incident light. In the example shown in the figure, the pixel substrate 120 is a back-illuminated type in which the substrate that is a base when the pixels 121 are formed is removed or thinned to receive incident light from the base substrate side.

  In addition, the pixel substrate 120 may include, for each light receiving element, a transistor that instructs reset, transfer, and selection, and an element that amplifies an output signal. The microlens 110 stacked on the pixel substrate 120 condenses incident light on each pixel 121 to improve incident efficiency.

  The AD conversion circuit board 130 includes a plurality of AD conversion circuits 131, a latch 132, and a changeover switch 133 that individually correspond to the pixels 121 of the pixel board 120. Accordingly, the AD conversion circuit board 130 outputs either a value obtained by discretizing each pixel value acquired from the pixel 121 of the pixel board 120 or a value held by the latch 132 to another board. The AD conversion circuit 131, the latch 132, and the changeover switch 133 operate at the timing when the timing trigger 210 received from the outside is received.

  The multiplication circuit board 140 includes a multiplication circuit 141 that individually corresponds to each pixel 121 of the pixel board 120. That is, the multiplication circuit substrate 140 includes the same number of multiplication circuits 141 as the plurality of pixels 121 provided on the pixel substrate 120. The multiplier circuit 141 is a digital multiplier and may be formed by, for example, a shift register. The multiplier circuit 141 acquires and holds a filter coefficient from the outside when executing the multiplication process.

  In other words, the multiplication circuit 141 can execute different multiplication processes according to the values of the filter coefficients read from the outside. FIG. 1 shows a case where the filter coefficient a is acquired as an example. The multiplication circuit 141 in the multiplication circuit board 140 also operates at the timing when the timing trigger 210 received from the outside is received.

  Note that the multiplication circuit substrate 140 may include one multiplication circuit 141 for each block including the plurality of pixels 121 of the pixel substrate 120. For example, if four pixels adjacent to each other in the two-dimensional direction of the pixel 121 are defined as one block, the multiplication circuit 141 may be connected to any of the four pixels 121 in the block. In this case, the multiplication circuit 141 sequentially performs multiplication processing on the outputs from the four pixels 121 in the block.

  The adder circuit 151 of the adder circuit board 150 adds and outputs the values acquired from the plurality of multiplier circuits 141 in the multiplier circuit board 140. The output of the addition circuit 151 can be output to the convolution calculation result addition board 160 in the lower layer in the drawing. The adder circuit 151 in the adder circuit board 150 operates at the timing when the timing trigger 210 received from the outside is received.

  The convolution operation result addition board 160 includes an addition circuit 161 with a latch B, a latch A 162, and a multiplexer 163. The adder circuit 161 with latch B, the latch A162, and the multiplexer 163 are connected to each other, the adder circuit 161 with latch B is added to the activation function arithmetic circuit board 170 through the through electrode 13, and the multiplexer 163 is added through the through electrode 13. Each is connected to the circuit board 150. The convolution calculation result addition board 160 adds a plurality of signals output from the addition circuit board 150 and outputs the result to the activation function calculation circuit board 170.

  The activation function arithmetic circuit board 170 has a number of activation function arithmetic circuits 171 corresponding to the adder circuits 151 of the adder circuit board 150. The activation function arithmetic circuit board 170 receives the output of the adder circuit board 150 and executes the activation function calculation, and then outputs it to the pooling circuit board 180. The activation function calculation circuit 171 in the activation function calculation circuit board 170 operates at the timing when the timing trigger 210 received from the outside is received.

  The pooling circuit 181 of the pooling circuit board 180 and the coupling circuit 191 of the coupling circuit board 190 sequentially process inputs from the previous stage. The output value of the combination circuit 191 on the combination circuit board 190 can be output as a feature amount to the outside of the feature extraction apparatus 100. The pooling circuit 181 and the coupling circuit 191 also operate at the timing when the timing trigger 210 received from the outside is received.

  In the feature extraction apparatus 100, the pixel 121, the AD conversion circuit 131, the latch 132, the changeover switch 133, the multiplication circuit 141, the addition circuit 151, the addition circuit 161 with the latch B, the activation function calculation circuit 171 and the pooling circuit 181. Each of the coupling circuit 191 and the like is controlled in operation timing by a control unit (not shown) by the supplied timing trigger 210. The control unit may be included in the feature extraction device 100, or another device equipped with the feature extraction device 100, for example, the control unit of the imaging device may be used. A feature extraction system is configured by the feature extraction device 100 that is a single element and the control unit.

  As described above, the feature extraction device 100 includes the multiplication circuit board 140, the addition circuit board 150, the convolution calculation result addition board 160, the activation function calculation circuit board 170, the pooling circuit board 180, and the combination involved in the image feature extraction processing. The circuit substrate 190 has a structure in which the pixel substrate 120 including the light receiving element is stacked. Thereby, since the pixel value can be directly processed and feature extraction can be performed, the processing for converting the image into data and the processing for transferring the stored image data become unnecessary, and the processing time can be shortened.

  In addition, hardware resources such as a storage device and a transfer device for image data can be omitted, which contributes to downsizing of the device including the feature extraction device 100. Furthermore, since the processing substrates corresponding to the pixels of the pixel substrate 120 are stacked, it is possible to prevent the processing speed from being lowered even if the number of pixels of the pixel substrate 120 is increased.

  Note that the feature extraction device 100 receives incident light from the plurality of pixels 121 arranged two-dimensionally, so that the two-dimensional luminance distribution information based on the generation of image data is obtained from the pixel substrate 120. It can also be acquired. Therefore, the feature extraction apparatus 100 can also be used as an image sensor.

  FIG. 2 is a flowchart of feature extraction processing executed in the feature extraction apparatus 100. As illustrated, in the feature extraction process, a convolution process S102, an activation function calculation process S103, a pooling process S104, and a combination process S105 are executed on the pixel value generated by the pixel value generation process S101, and extraction is performed. The obtained feature amount is output to the outside (step S106).

  Here, in the feature extraction process corresponding to deep learning, the pixel value generation process S101 and the feature amount output S106 are executed once for each feature extraction. However, the filter function reading, multiplication processing, and addition processing in the convolution processing S102 are repeatedly executed many times while changing the filter coefficient to be read. Furthermore, the processing results of the activation function calculation process S103 and the pooling process S104 after the convolution process are provided again to the convolution process S102, and the processes from the convolution process S102 to the pooling process S104 are further repeated. Note that only the activation function calculation process S103 is repeated, the convolution process S102 and the activation function calculation process S103 are repeated, or one or both of the convolution process S102 and the activation function calculation process S103 after the pixel value generation process S101. May be omitted and the pooling process S104 may be performed.

  FIG. 3 is a diagram for explaining the operation of the feature extraction apparatus 100 shown in FIG. In the feature extraction apparatus 100 shown in the figure, the function of the through electrode 13 for connecting adjacent substrates is highlighted by a thick line with hatching.

  In the feature extraction apparatus 100, each of the pixel 1, the pixel 2, and the pixel 3 in the pixel substrate 120 passes through the corresponding AD conversion circuit 131 and the changeover switch 133 in the AD conversion circuit substrate 130, as indicated by a thick line in the figure. Thus, the multiplication circuit board 140 is connected to the corresponding multiplication circuit 141.

  On the other hand, the multiplication circuit board 140 acquires the filter coefficients a, b, and c of the multiplication circuit 141 corresponding to each of the pixels 1 to 3. The pixel value output from the pixel 1 on the pixel substrate 120 is multiplied using the filter coefficient a by the multiplication circuit 141 corresponding to the pixel 1 in the multiplication circuit substrate 140 and then input to the addition circuit substrate 150 through the through electrode 13. The

  Similarly, the pixel value output from the pixel 2 on the pixel substrate 120 is multiplied using the filter coefficient b by the multiplication circuit 141 corresponding to the pixel 2 in the multiplication circuit substrate 140, and then added through the through electrode 13. Is input. Further, the pixel value output from the pixel 3 on the pixel substrate 120 is multiplied using the filter coefficient c by the multiplication circuit 141 corresponding to the pixel 3 in the multiplication circuit substrate 140, and then applied to the addition circuit substrate 150 through the through electrode 13. Entered.

  The adder circuit 151 in the adder circuit board 150 adds a plurality of input multiplication processing results and outputs the result to the convolution operation result adder board 160. In this way, processing similar to filtering such as smoothing is executed. However, a series of processing is executed as convolution processing by preliminarily learning filter coefficients in the multiplication circuit 141.

  The latch A 162 in the convolution calculation result addition board 160 holds the signal output from the addition circuit 151 via the multiplexer 163. After that, multiplication processing is performed by the multiplication circuit 141 in which a filter coefficient different from the previous one is set, and the multiplication processing results are added by the addition circuit 151. The addition circuit 161 with latch B then adds the addition processing result of the addition circuit 151, This is held through the multiplexer 163.

  The adding circuit 161 with the latch B reads the previous addition processing result held in the latch A 162 and adds it to the new addition processing result, and the addition result is again held by the latch A 162. By repeating this process a predetermined number of times, it is possible to add together a plurality of processing results obtained by performing a convolution operation with different filter coefficients for the same pixel group (pixels 1 to 3). That is, this corresponds to repetition of the convolution process S102 shown in FIG. Such a convolution operation is performed on all the pixels on the pixel substrate 120, that is, the entire input image.

In the feature extraction apparatus 100, the processing result by the convolution calculation result addition board 160 is input to the activation function calculation circuit 171 of the activation function calculation circuit board 170 through the through electrode 13. The activation function arithmetic circuit 171 converts information propagated from the convolution process to the pooling process. As such a function, as shown in the following Expression 1, a ReLU (Rectified Linear Unit) function that returns all input values less than 0 as 0 and returns positive values as they are can be exemplified.
f (x) = max (0, x) (Formula 1)

  The output of the activation function arithmetic circuit 171 is transmitted to the pooling circuit 181 of the pooling circuit board 180 through the through electrode 13. In the pooling circuit 181, subsampling is executed based on a predetermined condition. An example of the subsampling condition executed here is a Max Pooling method in which a maximum value is a representative value in a certain window size.

  The predetermined condition may be a process for collecting a plurality of outputs from the activation function arithmetic circuit 171, and may be an Average Pooling method for outputting an average value. Thus, the processes up to the convolution process (step S102), the activation function calculation process (step S103), and the pooling process (step S104) are executed, and a sub-sampling value generated from the pixel value is generated.

  FIG. 4 is a diagram for explaining another operation of the feature extraction apparatus 100 shown in FIG. In the feature extraction apparatus 100 shown in the figure, the through electrode involved in the operation to be described next is highlighted with a hatched thick line.

  As indicated by the bold lines in the figure, the feature extraction apparatus 100 can transmit and receive signals between substrates that are not adjacent to each other. Therefore, for example, the output of the pooling circuit board 180 can be stored in the latch 132 of the AD conversion circuit board 130 through the through electrode 14 and again subjected to multiplication processing in the multiplication circuit 141. As a result, the subsampled value is convolved again. Such repetition of the convolution process is executed a predetermined number of times, for example, 2000 times or more while changing the filter coefficients a, b, and c.

  Furthermore, signals processed up to a predetermined number of times between the AD conversion circuit board 130, the multiplication circuit board 140, the addition circuit board 150, the convolution calculation result addition board 160, the activation function calculation circuit board 170, and the pooling circuit board 180. Is made one-dimensional by the coupling circuit 191 on the coupling circuit board 190. Thereby, a value indicating a certain feature amount of the image is obtained for each component.

  FIG. 5 is a timing chart showing the operation of the feature extraction apparatus 100. The pulse in the figure means a signal supplied to each substrate as the timing trigger 210.

  As shown in the figure, each circuit in the feature extraction device 100 is supplied with a preparation timing trigger 210 in order from the upper layer in FIG. 1, thereby executing the processing shown in FIG. 2 in the order shown in FIG. In 100, since the multiplication circuit board 140 has the multiplication circuit 141 corresponding to each of the pixels 121, the convolution processing can be executed on each pixel value. Thereby, a feature amount with high determination accuracy can be extracted efficiently.

  The section P indicates a section in which the convolution process S102 is repeated. A region surrounded by a dotted line A indicates a section in which a convolution calculation is performed with a filter of the same size on a pixel group (for example, 3 × 3 pixels) composed of a plurality of pixels. Furthermore, a region surrounded by a dotted line B indicates a section where the same pixel group is subjected to a convolution operation by another filter.

  A section Q indicates a section in which the convolution process S102 to the pooling process S104 are repeated. Actually, in the calculation result addition process of the convolution process S102, the timing trigger of the addition circuit with the latch A or the latch B is supplied every time the signal from the addition circuit is input. Only the timing trigger of the final addition process is shown.

  Note that the coupling process in the coupling circuit 191 has a smaller processing load compared to the process on other substrates. Therefore, the feature extraction apparatus 100 does not have to provide the coupling circuit board 190 for the coupling circuit 191. In this case, the feature extraction apparatus 100 outputs a signal that is repeated a predetermined number of times and finally pooled by the pooling circuit 181 to the outside. Then, the coupling process is performed on the external substrate.

  FIG. 6 is a block diagram of an imaging apparatus 500 that includes the feature extraction apparatus 100. The imaging apparatus 500 includes a feature extraction device 100, a system control unit 501, a drive unit 502, a photometry unit 503, a work memory 504, a recording unit 505, a display unit 506, and a main power source 507.

  In addition, the imaging apparatus 500 includes a main optical system 520 that guides the subject light flux to the feature extraction apparatus 100. The main optical system 520 may be an exchangeable type that can be attached to and detached from the imaging apparatus 500.

  The main optical system 520 is composed of a plurality of optical lens groups, and forms an image of a subject light beam from the object field in the vicinity of the focal plane. In the figure, the main optical system 520 is representatively represented by a single virtual lens disposed in the vicinity of the pupil.

  The drive unit 502 is a control circuit that executes charge accumulation control such as timing control and region control of the feature extraction apparatus 100 in accordance with an instruction from the system control unit 501. The drive unit 502 performs, for example, a series of controls for causing the feature extraction device 100 to accumulate charges generated by photoelectric conversion of incident light and to output pixel values. In addition, the drive unit 502 supplies a timing trigger 210 to the feature extraction apparatus 100.

  The feature amount of the subject output from the feature extraction device 100 is passed to the determination unit 513 of the system control unit 501. As a result, the system control unit 501 executes determination processing for determining a subject. Note that the feature extraction apparatus 100 may transmit information including the luminance distribution of the subject from the pixel substrate 120 to the image processing unit 511 of the system control unit 501 to generate image data of the subject. The image processing unit 511 executes processing using the work memory 504 as a work space.

  The photometric unit 503 detects the luminance distribution of the subject prior to a series of sequences in which the feature extraction device 100 generates pixel values by photoelectrically converting incident light. The photometry unit 503 includes, for example, an AE sensor having about 1 million pixels. The calculation unit 512 of the system control unit 501 receives the output of the photometry unit 503 and calculates the luminance for each area of the scene.

  Further, the calculation unit 512 determines the shutter speed, aperture value, and ISO sensitivity according to the calculated luminance distribution. The photometry unit 503 may also be used in the feature extraction device 100. Note that the calculation unit 512 also executes various calculations necessary for operating the imaging apparatus 500.

  As described above, the imaging apparatus 500 has a determination function for determining a subject based on the feature amount acquired from the feature extraction apparatus 100. Here, the feature extraction apparatus 100 transmits the feature amount extracted by itself to the system control unit 501. Therefore, the system control unit 501 can determine the subject by acquiring the feature amount without incurring the load of the feature amount extraction process. Further, since the system control unit 501 receives the feature amount as the extraction result, an increase in the communication amount between the feature extraction apparatus 100 and the system control unit 501 is prevented.

  FIG. 7 is a schematic cross-sectional view of another feature extraction apparatus 101. The feature extraction apparatus 101 has the same structure as the feature extraction apparatus 100 shown in FIG. Therefore, common elements are denoted by the same reference numerals, and redundant description is omitted.

  The feature extraction apparatus 101 has a different structure from the feature extraction apparatus 100 in that a plurality of multiplication circuit boards 1400, 1401, 1402,... 140n are provided between the AD conversion circuit board 130 and the addition circuit board 150. The plurality of multiplication circuit boards 1400, 1401, 1402,... 140n have a plurality of multiplication circuits 141 that individually correspond to the plurality of pixels 121 of the pixel board 120, similarly to the multiplication circuit board 1400 of the feature extraction device 100. The operation is performed at the timing when the timing trigger 210 received from the outside is received.

  The multiplier circuit 141 acquires and holds a filter coefficient from the outside when executing the multiplication process. In addition, the multiplication circuit 141 can perform multiplication processing while holding filter coefficients having different values for each substrate. Therefore, the multiplication circuit boards 1400 to 140n can perform multiplication processing under different conditions.

  FIG. 8 is a schematic cross-sectional view for explaining the operation of the feature extraction apparatus 101. As shown by the hatched thick line in the figure, the pixel value output from the pixel 1 on the pixel substrate 120 is multiplied by the multiplication circuit 141 formed on the multiplication circuit board 1400 in the upper stage in the figure, and then the through electrode 13 is input to the adder circuit board 150. The pixel value output from the pixel 2 on the pixel substrate 120 is multiplied by the multiplication circuit 141 formed on the multiplication circuit substrate 1401 in the upper part of the drawing, and then input to the addition circuit substrate 150 through the through electrode 13. Further, the pixel value output from the pixel 3 on the pixel substrate 120 is multiplied by the multiplication circuit 141 formed on the multiplication circuit substrate 1402 in the upper part of the drawing, and then input to the addition circuit substrate 150 through the through electrode 13.

  The filter coefficient a1 in each of the plurality of multiplication circuits 141 provided on the multiplication circuit board 1400 has the same value between the multiplication circuits 141 provided on the same multiplication circuit board 1400. Similarly, in each of the other multiplication circuit boards 1401, 1402,... 140n, the plurality of multiplication circuits 141 in each board have common filter coefficients b1, c1,. Therefore, when executing the convolution process, the filter coefficients a1, b1, c1,..., N1 can be selected by selecting the substrate on which the multiplication process is executed. That is, in the example shown in FIG. 8, the pixel value output from the pixel 1 is multiplied by the filter coefficient a1, the pixel value output from the pixel 2 is multiplied by the filter coefficient b1, and the pixel value output from the pixel 3 is filtered. Multiplication processing is performed by the coefficient c1.

  FIG. 9 is a partial timing chart of the feature extraction apparatus 101. In the feature extraction apparatus 101, multiplication processing is performed by different multiplication circuit boards 1400 to 140 n for each pixel 121. Therefore, after first obtaining all the filter coefficients, the multiplication process executed many times in the convolution process corresponding to one filtering for a certain pixel group is shown surrounded by dotted lines A and B in the figure. In addition, it can be processed in parallel and simultaneously.

  Therefore, it is not necessary to acquire and set the filter coefficient every time the multiplication process is performed. Therefore, the time for acquiring and setting the filter coefficient is shortened, and the throughput of the entire feature extraction apparatus 101 can be improved. Such a feature extraction apparatus 101 can be used as an image sensor with a feature extraction function by being incorporated in the imaging apparatus 500 instead of the feature extraction apparatus 100, for example.

  The feature extraction apparatus described in this embodiment includes a multiplication circuit 141, an addition circuit, and a multiplication circuit board 140, an addition circuit board 150, an activation function arithmetic circuit board 170, a pooling circuit board 180, and a coupling circuit board 190, respectively. 151, an activation function operation circuit 171, a pooling circuit 181, and a coupling circuit 191 are arranged. However, the feature extraction apparatus does not necessarily need to be provided with only one circuit for one substrate. That is, a plurality of circuits may be arranged on one substrate, or one circuit may be arranged across a plurality of substrates.

  In addition, the stacking order of the plurality of substrates in the feature extraction device is not limited to the above example as long as it can be arranged using the through electrodes. Furthermore, the convolution calculation result adding board 160 is not limited to the above example as long as it can add and hold a plurality of values.

  In addition, the feature extraction apparatus described in the present embodiment realizes feature extraction using a convolutional neural network that performs convolution processing in a multilayer neural network using a single sensor having a stacked structure. It is also possible to realize a recursive convolutional neural network by using the same value for the filter coefficient of the convolution process performed a plurality of times. Note that there is no limitation to these methods as long as the feature extraction can be realized by the feature extraction apparatus according to the present embodiment.

  As mentioned above, although this invention was demonstrated using embodiment, the technical scope of this invention is not limited to the range as described in the said embodiment. It will be apparent to those skilled in the art that various modifications or improvements can be added to the above-described embodiment. It is apparent from the scope of the claims that the embodiments added with such changes or improvements can be included in the technical scope of the present invention.

  The order of execution of each process such as operations, procedures, steps, and stages in the apparatus, system, program, and method shown in the claims, the description, and the drawings is particularly “before” or “prior to”. It should be noted that the output can be realized in any order unless the output of the previous process is used in the subsequent process. Regarding the operation flow in the claims, the description, and the drawings, even if it is described using “first”, “next”, etc. for convenience, it means that it is essential to carry out in this order. It is not a thing.

DESCRIPTION OF SYMBOLS 11 Ground substrate, 12 Wiring layer, 13, 14 Through electrode, 100, 101 Feature extraction device, 110 Micro lens, 120 Pixel substrate, 121 pixel, 130 AD conversion circuit board, 131 AD conversion circuit, 132 Latch, 133 Changeover switch, 140, 1400, 1401, 1402, 140n multiplication circuit board, 141 multiplication circuit, 150 addition circuit board, 151 addition circuit, 160 convolution operation result addition board, 161 addition circuit with latch B, 162 latch A, 163 multiplexer, 170 activation Function arithmetic circuit board, 171 Activation function arithmetic circuit, 180 Pooling circuit board, 181 Pooling circuit, 190 Coupling circuit board, 191 Coupling circuit, 210 Timing trigger, 500 Imaging device, 501 System control unit, 502 Drive unit, 503 Photometry unit, 504 Work memory, 505 Recording unit, 506 Display unit, 507 Main power source, 511 Image processing unit, 512 Calculation unit, 513 Determination unit, 520 Main optical system

Claims (10)

A light receiving substrate in which a plurality of light receiving elements for photoelectrically converting received light are two-dimensionally arranged;
Comprising one or more other substrates stacked on the light receiving substrate;
The other substrate is
A plurality of multiplication circuits provided corresponding to each of the light receiving elements or each block composed of the plurality of light receiving elements, and a convolution operation for signals output from the plurality of light receiving elements using the plurality of multiplication circuits A convolution processing unit,
A pooling processing unit that samples a signal output from the convolution processing unit based on a predetermined condition;
And a connection wiring for passing the sampled signal to the plurality of multiplication circuits. The convolution processing unit and the pooling processing unit are respectively provided on different substrates among the plurality of other substrates,
The feature extraction element according to claim 1, wherein the connection wiring includes a through electrode that connects the different substrates. A plurality of multiplier circuit boards stacked on each other as the other board;
A multiplication circuit,
3. The feature extraction element according to claim 1, wherein the plurality of multiplication circuits provided for each of the light receiving elements or the blocks are provided on different multiplication circuit boards.   The feature extraction element according to claim 3, wherein different filter coefficients for convolution calculation are set for the plurality of multiplication circuit boards. The convolution processing unit includes an addition circuit that adds outputs from the plurality of multiplication circuits provided corresponding to each of the plurality of light receiving elements,
The feature extraction element according to claim 1, wherein the adder circuit is provided on an adder circuit board that is the other board. A latch circuit for holding a signal from the light receiving element; a readout circuit for reading a signal from the light receiving element;
The readout circuit is provided on a readout substrate which is the other substrate,
The feature extraction element according to claim 1, wherein the connection wiring connects the pooling processing unit and the latch circuit.   The said other board | substrate has an activation function calculating process part which inputs the signal by which the said convolution calculation was input into an activation function, and outputs the output signal to the said pooling process part. The feature extraction element described in 1.   The feature extraction element according to claim 1, wherein signals output from the plurality of light receiving elements are output to the outside. The feature extraction element according to any one of claims 1 to 8,
A convolution operation of the convolution processing unit, and a control unit that controls to repeat sampling of the pooling processing unit,
The said control part is a feature extraction system which controls the said convolution process part so that a predetermined filter coefficient may be used, respectively, when performing the said convolution operation repeatedly. The feature extraction element according to any one of claims 1 to 8,
A determination unit that determines an imaging target based on a feature amount extracted based on an output from the pooling processing unit;
A determination apparatus comprising:

Images