CN103973991B - A kind of automatic explosion method judging light scene based on B P neutral net - Google Patents

Abstract

The invention discloses an automatic exposure method for judging a lighting scene based on a B-P neural network, comprising the following steps: S1 obtains an original image through a video acquisition system; S2 divides the original image into multiple areas; S3 calculates the image brightness of each area The average value is used to obtain the brightness vector; S4 designs the B‑P neural network, and uses the brightness vector as the input of the neural network to judge the lighting scene; S5 calculates the ideal brightness of the image according to the judgment result of the neural network; S6 uses the actual brightness of the original image The deviation between the brightness and the ideal brightness is used as the initial input of the PID algorithm, and the PID algorithm is used to obtain the ideal controllable quantity corresponding to the ideal brightness; S7 obtains the exposure time t and the analog gain coefficient g according to the ideal controllable quantity, and transmits t and g to the video Automatic exposure can be realized in the sensor of the acquisition system. The method can accurately judge the lighting scene, has a simple algorithm, and can be widely used.

Description

An automatic exposure method for judging lighting scenes based on B-P neural network

technical field

The invention belongs to the technical field of video image processing, and in particular relates to an automatic exposure method for judging lighting scenes based on a B-P neural network.

Background technique

Automatic exposure refers to the method of automatically adjusting the aperture, shutter, signal gain, etc., so as to obtain a clear image with a color close to the real object. Since the birth of the point-and-shoot camera, technicians in related fields have been studying automatic exposure methods. The early exposure methods used analog exposure, which was fast and did not require a lot of calculations, but could not perform scene analysis. Although exposure compensation could be performed, it was still difficult to obtain ideal image brightness, and its mechanical and electronic structure was complex and poor in stability.

After the birth of digital image technology, the electronic exposure method gradually replaced the analog exposure method. The electronic exposure method only needs to analyze the digital image to obtain the light intensity, and then adjust the aperture and shutter to perform automatic exposure. With embedded devices such as single-chip microcomputers , DSP, FPGA and other hardware programming equipment are gradually used in digital image acquisition. Image acquisition equipment has some basic image analysis functions. In order to obtain clearer images with colors closer to real objects, technicians introduce simple scene analysis techniques into automatic exposure. In the method, a dynamic ideal brightness value is adopted to adapt to various external lighting conditions, such as backlight and front light, so that the exposure method is performed based on the lighting scene, that is, an exposure method for judging the lighting scene is produced.

The image analysis technology based on image entropy to set the ideal exposure value is an exposure method based on lighting scenes, so that the image subject has a good exposure no matter what the lighting conditions are, but the algorithm of image entropy is relatively complicated, making it applicable to many areas. Limited and not widely applicable. There is also an automatic exposure method based on face recognition and fuzzy logic, which is relatively complicated to implement and cannot be widely used.

Contents of the invention

In view of the above problems, the purpose of the present invention is to provide an automatic exposure method for judging lighting scenes based on B-P neural network, which is accurate in judging lighting scenes, has a simple algorithm, and can properly expose image subjects under backlight and light conditions. It meets the automatic exposure requirements of the video processing front end under the network camera application.

In order to achieve the above object, the invention provides a kind of automatic exposure method based on B-P neural network judgment lighting scene, it is characterized in that, comprises the following steps:

S1: Obtain the original image by the video acquisition system;

S2: divide the original image into a plurality of regions according to the degree of attention to different regions of the picture, and number the regions according to the degree of attention;

S3: Find the average brightness of the image in each region for the divided regions, and obtain a brightness vector in conjunction with the region position number and the average brightness of the image in the region;

S4: according to the quantity design B-P neural network of described original image region division, described luminance vector is used as the input of neural network, the lighting scene is judged, output judgment result;

S5: Calculate the ideal brightness of the image according to the judgment result of the neural network;

S6: Using the deviation between the actual brightness of the original image and the ideal brightness in S5 as the initial input of the PID algorithm, using the PID algorithm to obtain the ideal controllable amount required by the ideal brightness;

S7: Obtain the ideal exposure time t and ideal analog gain coefficient g according to the ideal controllable amount obtained in S6, and transmit the obtained ideal exposure time t and ideal analog gain coefficient g to the sensor of the video acquisition system, which can be realized automatic exposure.

Further, the number of input layer variables of the B-P neural network in the step S4 is the number of divisions of the original image area, and the designed B-P neural network output formula is as follows:

y=f(X), y=0,1

X is the vector representation of the input data of the neural network, y is the output of the neural network, y=0 or 1, that is, there are two types of judgment results output by the B-P neural network, 0 represents normal lighting, and 1 represents special lighting conditions. Such a design can quickly determine the structure of the neural network and simplify its structure, which is beneficial to its learning and training process.

Further, in the step S5, when the output of the neural network is 0, the ideal brightness takes half of the maximum component of the brightness vector;

When the neural network output is 1, the formula for calculating the ideal brightness is as follows:

In the formula, yz is the rating brightness of the image subject, X is the brightness vector of the image, W is the rating brightness weight vector, _{y p} is the evaluation brightness of the picture, η is the brightness error coefficient of the image subject, l is the correction coefficient, which takes The value range is between 0 and 1, y _mid is one-half of the largest component of the brightness vector, that is, the middle brightness of the image, y _l is the calculated ideal brightness, W _i is the ith component of the rating brightness weight vector W, T stands for transpose.

Among them, the rating brightness weight vector W represents the degree of attention to different areas of the image, and its components are the same as the number of components in the brightness vector X and correspond to each other. The higher the degree of attention to a certain area, the value of the component corresponding to the area in the weight vector bigger.

Further, the evaluation brightness y _p of the picture is calculated using the following formula:

y _p =X*W _z ^T

Among them, y _p is the evaluation brightness of the picture, X is the brightness vector of the image, and W _z is the evaluation brightness weight vector, which is designed according to experience.

Among them, the evaluation brightness weight vector _Wz represents the degree of attention in different areas of the image, and its components are the same as the number of components in the brightness vector X and correspond to each other. The higher the degree of attention in a certain area, the evaluation brightness weight vector corresponds to the area The larger the component value is. The difference between the evaluation brightness weight vector W _z and the rating brightness weight W is that the values of each component are different.

Further, in the step S6, the controllable quantity is expressed by the following formula:

lnp=lntg

In the formula, lnp is the controllable quantity, p=tg, t is the automatic exposure time, and g is the analog gain coefficient.

Further, in the step S7, the exposure time t and the analog gain coefficient g are obtained by querying the exposure gain table. There are many ways to obtain the exposure time t and the analog gain coefficient g according to the controllable quantity lnp, such as look-up table method, iterative method, or other control algorithms, and querying the exposure gain table is one of the more convenient and fast methods.

In the method of the present invention, the B-P neural network is designed to judge the lighting scene, and the B-P neural network is trained and learned, so that the B-P neural network has certain flexibility and intelligence, and the actual image brightness vector is used as the input of the neural network, thereby realizing accurate lighting scene Judgment: Only after accurate judgment is made on the lighting scene, can appropriate exposure be taken based on the lighting scene. The method of the present invention has the advantages of accurate judgment on the lighting scene, simple algorithm, strong versatility, and can be widely used.

Description of drawings

FIG. 1 is a flowchart of an automatic exposure method according to an embodiment of the present invention;

Fig. 2 is the division and numbering of the original image area in the embodiment of the present invention;

Fig. 3 is the B-P neural network structural diagram of design in the embodiment of the present invention;

Fig. 4 is the neuron model of B-P neural network in the present embodiment;

FIG. 5 is an effect diagram of brightness correction based on ideal brightness in an embodiment of the present invention;

Fig. 6 is the system block diagram that adopts PID controller to obtain ideal controllable quantity in the embodiment of the present invention;

Fig. 7 is the final exposure gain table of camera ov9712 in the embodiment of the present invention;

FIG. 8 is a comparison diagram of effects obtained by applying the automatic exposure method in the embodiment of the present invention in different lighting scenarios.

detailed description

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

Fig. 1 is the flow chart of the automatic exposure method of embodiment in the present invention, and this method comprises:

S1: The original image is obtained through the video acquisition system. The light containing the scene signal is focused on the CMOS image sensor after being focused by the lens. The CMOS image sensor generally includes three parts: COMS photosensitive element, analog gain coefficient device, and A/D converter. , the light irradiated on the COMS sensor is converted into an analog electrical signal through the photosensitive element after a certain period of exposure, the analog gain coefficient device amplifies the electrical signal, and finally converts the electrical signal into a digital signal through the A/D converter, and converts the digital signal into a digital signal. The signal is output to the digital video acquisition controller.

S2: Divide the original image into a plurality of regions according to the degree of attention to different regions of the picture, and number the regions in turn according to the degree of attention, taking into account the characteristics of the picture and people's attention to the picture in the case of forward light or backlight Degree and experience, divide the picture into 13 areas, think that the middle position of the picture is the area with the most attention, its number is smaller, and the surrounding area of the picture is the area with less attention, its number is larger, according to the area of attention Degree, the divided areas are numbered 1 to 13 in sequence, and the divided areas and numbers of the image are shown in Figure 2.

S3: Calculating the average brightness of the image in each region for the divided regions, combining the position number of the region and the average brightness of the image in the region to obtain a brightness vector. For the 13 regions divided by S2, calculate the average value of the image brightness in each region, let it be x _i , i is the position number of the region, i=1,2...13, then the brightness vector X=(x ₁ ,x ₂ ,...,x ₁₃ ).

S4: Design the B-P neural network according to the number of divisions of the original image area, use the brightness vector as the input of the neural network, judge the lighting scene, and output the judgment result. The number of input layer variables of the B-P neural network is the number of divisions of the original image area, which is 13 in the present embodiment, and the output formula of the B-P neural network of design is as follows:

y=f(X), y=0,1

X is the vector representation of the input data of the neural network, y is the output of the neural network, y=0 or 1, that is, there are two types of judgment results output by the B-P neural network, 0 represents normal lighting, and 1 represents special lighting conditions.

Determine the number of input layer variables of the neural network to be 13, and the number of output neurons to be 1. Now it is only necessary to determine the number of layers in the middle hidden layer and the number of neurons in each layer. The structure of the B-P neural network is can be determined.

The B-P neural network needs to be trained and learned before it can have certain flexibility and intelligence, and then it can make accurate judgments on lighting scenes. There are many training and learning algorithms for the B-P neural network. In this embodiment, the error backpropagation learning algorithm is preferred, but the present invention does not limit the training and learning algorithm for the B-P neural network.

The middle hidden layer can be one or more layers. When the middle hidden layer is multi-layered, the error backpropagation learning algorithm cannot obtain a better training and learning effect. When the middle hidden layer is one layer, the structure of the B-P neural network is simple. The error backpropagation learning algorithm is mature, so the middle hidden layer is used as one layer, but the present invention does not limit the number of layers of the middle hidden layer of the neural network.

The number of neurons in the middle hidden layer is mostly set by experience, and the general reference calculation formula is as follows:

l<n-1

l=log ₂ n

Among them, l is the number of neurons in the hidden layer, n is the number of neurons in the input layer, m is the number of output neurons, a is a constant, and its range is 0-10. According to experience, if the constant a is 3, the middle hidden The number of layer neurons is 6, and the structure of the neural network designed in this embodiment is obtained as shown in FIG. 3 .

Next, it is necessary to determine the parameters of each neuron. Neuron parameters include input weights, neuron offset values, and transfer functions. The neuron model is shown in Figure 4. Assuming that Figure 4 shows the jth neuron, the neuron The input of each element is p _j1 to p _jn , the offset function is b _j , the transfer function is f _j , and the weights are w _j1 to w _jn .

The transfer function is generally a simoid function or a logistic function. When the transfer function is a simoid function, the B-P neural network can efficiently complete the training process by using the steepest descent method, so all neurons use the simoid function as the transfer function.

The process of determining the input weight and offset value of each neuron is the training and learning process of the neural network. The B-P neural network adopts the learning process of the error back propagation learning algorithm. This algorithm is relatively mature and has been integrated into the Matlab toolbox. Among them, it can be implemented by programming in Matlab, or directly use the Matlab toolbox. In this embodiment, the Matlab toolbox is directly called to complete the learning process.

The B-P neural network adopts the learning process of the error back propagation learning algorithm. The sample data set used in the learning process has a great influence on the speed of training and learning and the accuracy of the final weight and offset value. For the parameters of each neuron in the network, it is necessary to correctly select learning sampling points. Considering that the sampling points must cover all possible parameters and the application objects of the neural network, the sample set data is designed in Table 1 in this embodiment.

Table 1 Neural network training sample set situation

lighting conditions Number of samples y output Aiming at light 1000 1 backlight condition 1000 1 Indoor normal light 1000 0 outdoor is shining 1000 0

Convert the image luminance of the image samples designed in Table 1 into a luminance vector as the input of the neural network, that is, calculate the luminance vector X=(x ₁ , x ₂ ... x ₁₃ ) of each sample, and use each component of the luminance vector as BP input to the neural network.

Substituting the sample data in Table 1 into the error backpropagation learning algorithm module of the Matlab toolbox, the B-P neural network of this embodiment is trained and learned, and the input weights and offset values of each neuron can be obtained through training and learning.

Table 2 shows the input weights from the input layer to the hidden layer obtained through training and learning, and the offset values of each neuron in the hidden layer.

Table 3 shows the input weights from the hidden layer to the output layer and the offset value of the neurons in the output layer.

Table 1 Neural network output layer to hidden layer information

neuron 1 neuron 2 neuron 3 neuron 4 neuron 4 Neuron 6 _x1 0.1276 -0.3372 -0.0845 -0.1833 -0.8470 -0.0714 _x2 0.5879 -1.5609 -0.6589 0.5194 0.5053 0.0397

_x3 -0.1543 -0.2152 0.0366 -0.9620 0.4315 0.2737 _x4 0.0622 -0.4902 0.7891 -0.8363 0.0391 0.3279 _x5 1.0354 -0.6431 -0.3181 0.0359 -0.8763 0.1632 _x6 -0.5359 -0.0690 0.6762 -0.6810 -0.0480 -0.4427 _x7 -0.1343 -0.7522 0.0114 0.1146 -0.2919 -0.7228 _x8 0.5234 0.2893 0.8203 -0.1602 0.8097 0.3045 _x9 -0.4793 0.0864 0.6060 -0.7993 -0.0423 0.0958 x ₁₀ 1.0091 -1.1566 0.1749 0.8206 0.2858 -0.0081 x ₁₁ 0.1433 0.0517 0.0961 -0.6016 0.4484 0.1029 x ₁₂ 0.4519 0.8756 1.1838 -0.8442 -0.5452 -0.7440 x ₁₃ -0.1887 -0.5859 0.8504 -0.5049 0.1127 0.5854 offset b -1.4628 1.7321 1.0325 -0.5727 -0.9139 -2.0257

Table 2 Neural network hidden layer to output layer information

Apply the neuron parameters obtained in Table 2 and Table 3 to the B-P neural network in this embodiment, and randomly select backlit photos, photos against the light, photos under normal lighting conditions, and photos where the scene cannot be judged by applying image entropy as experiments The sample is used to verify the accuracy of the B-P neural network designed in this embodiment in judging the lighting scene. The results show that the B-P neural network in this embodiment is accurate in judging the lighting scene of the picture, and it can judge the lighting scene of the picture where the application of image entropy to judge the scene fails.

S5: Calculate the ideal brightness of the image according to the judgment result of the neural network. When the output of the neural network is 0, the ideal brightness takes half of the maximum component of the brightness vector; when the output of the neural network is 1, the formula for calculating the ideal brightness is as follows :

In the formula, yz is the rating brightness of the image subject, X is the brightness vector of the image, W is the rating brightness weight vector, _{y p} is the evaluation brightness of the picture, η is the brightness error coefficient of the image subject, l is the correction coefficient, which takes The value range is between 0 and 1, y _mid is one-half of the largest component of the brightness vector, that is, the middle brightness of the image, y _l is the calculated ideal brightness, W _i is the ith component of the rating brightness weight vector W, T stands for transpose.

The evaluation brightness y _p of the picture is calculated by the following formula:

y _p =X*W _z ^T

Wherein, y _p is the evaluation brightness of the picture, X is the brightness vector of the image, X=(x ₁ , x ₂ ... x ₁₃ ), W _z is the evaluation brightness weight vector, which is designed according to experience.

Among them, the rating brightness weight vector W and the evaluation brightness weight vector W _z both represent the degree of attention to different areas of the image, and their components are the same as the number of components in the brightness vector X and correspond to each other. The larger the value of the component in the weight vector corresponding to the region. The evaluation brightness weight vector W _z is different from the magnitude of each component value of the rating brightness weight vector W. According to experience, W _z =[4,1,1,1,1,0,0,0,0,0,0, 0,0], W=[4,2,2,2,2,1,1,1,1,1,1,1,1], the correction coefficient is l=0.5.

Figure 5 is a comparison effect diagram of correcting photos taken under backlight and under light scenes according to the ideal brightness. Figure 5(a) and Figure 5(c) are the original pictures taken under backlight and under light scenes respectively. Figure 5( b) and Figure 5(d) are pictures corrected according to the ideal brightness. The brightness of the corrected picture is more suitable for human observation than the brightness of the picture before correction, illustrating the formula for calculating the ideal brightness and the rating brightness weight vector in the embodiment of the present invention The design of W and the design of evaluation brightness weight vector W _z are reasonable.

S6: Use the deviation between the actual brightness of the original image and the ideal brightness in S5 as the initial input of the PID algorithm, and use the PID algorithm to obtain the ideal controllable amount required by the ideal brightness.

First, the controllable quantity is expressed by the following formula:

lnp=lntg

In the formula, lnp is the controllable quantity, p=tg, t is the automatic exposure time, and g is the analog gain coefficient.

Among them, the process of obtaining the controllable quantity formula lnp=lntg is:

Under the premise of considering DC offset and other factors, the traditional expression formula of image brightness y after actual exposure is:

y＝πkltgr ² +c ₁ tg+c ₂

Among them, k is the proportional coefficient, l is the illuminance from the external environment to the lens, which is related to many factors such as lighting conditions, surface reflectance of the object to be photographed, and shooting angle, etc., r is the aperture radius of the image acquisition device, and the aperture radius is generally a fixed value , t is the exposure time, g is the analog gain coefficient, both t and g can be changed in size, c ₁ and c ₂ are constants related to the structure of the image acquisition system, which are fixed values.

The DC offset generated in the photosensitive process and the analog gain coefficient process is removed by black level correction, then c ₁ tg+c ₂ =0, and the actual image brightness y at this time is expressed by the following formula,

y=πkltgr ²

Simplify the above formula of actual image brightness, the simplification method is as follows:

The aperture radius r is a constant, the external scene illumination l is an uncontrollable quantity, the exposure time t and the analog gain coefficient g are controllable quantities, then p is a controllable quantity, q is an uncontrollable quantity, and k' is a constant.

At this time, the relationship between the actual image brightness y and the controllable quantity p and the uncontrollable quantity q is nonlinear, and the logarithm of both sides is taken to obtain the following formula,

lny=lnk'+lnp+lnq

In the formula, the relationship between lny and lnp and lnq is linear, and lnp is a controllable quantity, and lnq is a non-controllable quantity.

The effect of the above simplification is: separate the factors that affect the actual image brightness y, that is, the image brightness y is only related to p, p=tg, that is, the brightness of the actual image can be adjusted by adjusting the exposure time t and the analog gain coefficient g.

Express the ideal brightness as the brightness of the output image under the ideal controllable amount, obtain the ideal controllable amount according to the ideal brightness of the image calculated in step S5, and further obtain the exposure time t and the analog gain coefficient g required for the ideal brightness . In this embodiment, the ideal controllable quantity is obtained by using the PID algorithm, and the initial input of the PID algorithm is the deviation between the actual brightness of the original image and the ideal brightness of the image.

Figure 6 is a system block diagram of the ideal controllable quantity obtained by using the PID controller, where y _e in lny _e is the ideal brightness, y _out in lny _out is a certain brightness value close to the ideal brightness output after the PID algorithm, lny _In , y _in is the image brightness input to the digital video capture device after PID calculation, lnp is a controllable quantity, lnq+lnk' exists as system interference, and G(s) is before the digital video capture device to automatic exposure control Transfer function, H(s) is the feedback transfer function for seeking error. The deviation between the ideal luminance y _e and y _out is used as the input of the PID algorithm for many times, and the calculation is repeated in this way, that is, by adjusting the proportional coefficient k _p , the integral coefficient k _i , and the differential coefficient k _d in each calculation, the controllable quantity is closer to the ideal value, that is, even if lny _out is closer to lny _e , when |lny _out -lny _e |<h _e , the controllable quantity lnp at this time is considered to be an ideal controllable quantity, and the PID calculation process ends. It is proved by experiments that when The effect is better when the value of brightness threshold he is _ln1.1 .

S7: According to the ideal controllable amount obtained in S6, the ideal exposure time t and the ideal analog gain coefficient g are obtained by using the look-up table method, and the obtained ideal exposure time t and the ideal analog gain coefficient g are transmitted to the sensor of the video acquisition system , to achieve automatic exposure.

There are many ways to obtain the exposure time t and the analog gain coefficient g according to the controllable quantity lnp, such as look-up table method, iterative method, or other control algorithms, and querying the exposure gain table is a relatively simple one. The exposure gain tables of different types of lenses are different, that is, the functional relationship between the exposure time t and the analog gain coefficient g of different types of lenses is different. In the embodiment of the present invention, the camera ov9712 produced by ov company is used to illustrate the method of obtaining the exposure gain table.

The ov9712 image sensor is a COMS image sensor, and the gain coefficient adjustment range is 1-31. According to the characteristics of the COMS image sensor, in order to ensure that there are no streaks and flickering phenomena under the light, the exposure time and gain coefficient need to meet the requirements:

f _v =2f/n

t=m/(2f)

t<=1/f _v

1≤g≤31

U=lntg

In the formula, U is the function P in the controllable quantity lnp, m and n are proportional coefficients, adjusting the size of m can change the exposure time, adjusting the size of n, can adjust the size of the video sampling frame rate f _v , f is the AC frequency, f _v is the video sampling frame rate, t is the exposure time, and g is the analog gain coefficient. Wherein, AC frequency f=50hz, 1≤g≤31.

Because under different illuminance conditions, images with better brightness can be obtained by adjusting the exposure time, that is, under different illuminance conditions, exposure time adjustment has a greater impact on image brightness. Therefore, when determining the exposure gain table, first determine the exposure time table. Then the calculation gain table is adjusted according to the restriction conditions to obtain the final exposure gain table.

The specific method is, in the case of low illuminance, that is, when the illuminance is less than 10lux, take n=8, reduce the sampling rate to 12.5hz, and set the exposure time to 80ms, and an image with appropriate brightness can be obtained. In the case of strong light, that is, the illuminance is greater than At 2000lux (normal light illumination is about 500lux), regardless of the exposure time, no matter how short the exposure time is, overexposure will occur. At this time, the exposure time can be set to be greater than 10ms or less than 10ms. In order of thought, first determine the exposure schedule of the camera, and then adjust and calculate the gain table of the camera according to the restriction conditions, and obtain the final exposure gain table as shown in Figure 7.

Fig. 8 is a picture obtained by applying the automatic exposure method in the embodiment of the present invention in different lighting scenes. Figure 8(a) is an outdoor picture taken by the automatic exposure method in the embodiment of the present invention, the picture is suitable for human observation, and there is no exposure overflow; Figure 8(b) is the automatic exposure in the embodiment of the present invention The indoor picture taken by the method is suitable for human observation, and there is no underexposure or overexposure; Figure 8(c) and (d) are photos taken of the same object under the same conditions, and Figure 8(c) is the backlight Under the circumstances, there is no picture that should be taken by the automatic exposure method in the embodiment of the present invention. The brightness of the picture is relatively dark, the details cannot be recognized, and it is not suitable for human observation. Figure 8(d) is a picture taken by the automatic exposure method in the embodiment of the present invention , compared with Figure 8(c), its brightness is improved, and details can be recognized; Figure 8(e) and (f) are pictures taken of the same object under the same scene conditions, and Figure 8(e) is under the condition of light There is no picture taken by the automatic exposure method in the embodiment of the present invention. The brightness of the picture is relatively bright, the details cannot be recognized, and it is not suitable for human observation. Figure 8(f) is a picture taken by the automatic exposure method in the embodiment of the present invention. Compared with Figure 8(c), its brightness is reduced, and details can be recognized, which is suitable for human observation. The automatic exposure method in the embodiment of the present invention is applicable to various lighting scenes, and can obtain better image effects.

The automatic exposure method of the present invention can meet the front-end automatic exposure requirements of video processing under the application of network cameras, and the analysis results can be applied to important target segmentation, object recognition, adaptive video compression, content-sensitive video scaling, image retrieval and security monitoring, Applications such as military guards.

The above embodiments are only used to illustrate the present invention, but not to limit the present invention. Those of ordinary skill in the relevant technical field can make various changes and improvements without departing from the spirit and scope of the present invention. Therefore, all Equivalent technical solutions also belong to the category of the present invention, and the scope of patent protection of the present invention should be defined by the claims.

Claims (6)

1. a kind of the automatic explosion method of light scene is judged it is characterised in that comprising the following steps based on B-P neutral net：

S1:Original image is obtained by video acquisition system；

S2:According to the degree of concern of picture zones of different, described original image is divided into multiple regions, and according to being concerned The size of degree gives zone number successively；

S3:The brightness of image meansigma methodss in each region are asked to the region of described division, in conjunction with this regional location numbering and its region The mean flow rate of interior image, obtains luminance vector；

S4:According to the quantitative design B-P neutral net of described original image region division, using described luminance vector as nerve net The input of network, judges to light scene, exports judged result；

S5：According to the judged result of described neutral net, calculate the desired level of image；

S6：Using the deviation of desired level in the intrinsic brilliance of described original image and described S5 as pid algorithm initial input, Obtain the preferable controlled amounts required for desired level using pid algorithm；

S7：According to obtaining preferable controlled amounts in described S6, obtain ideal exposure time t and preferable analog gain coefficient g, will obtain Ideal exposure time t and preferable analog gain coefficient g be transferred in the sensor of video acquisition system, you can realize automatically exposing Light.

2. a kind of automatic explosion method judging light scene based on B-P neutral net according to claim 1, its feature It is, in described step S4, the input layer variable number of B-P neutral net is the number of original image region division, the B- of design P neutral net output formula is as follows：

Y=f (X), y=0,1

X is the vector representation of neutral net input data, and y is the output of neutral net, y=0 or 1, i.e. described B-P nerve net The result of determination of network output is two kinds, and 0 represents normal illumination, and 1 represents special light conditions.

3. a kind of automatic explosion method judging light scene based on B-P neutral net according to claim 1 and 2, its It is characterised by, in described step S5, when neutral net is output as 0, desired level takes two points of the largest component of luminance vector One of；

When neutral net is output as 1, the formula calculating desired level is as follows：

$\left\{\begin{array}{c}{y}_z=X*W^T/sum\left(W_i\right)\\ \mathrm{\&eta;}=y_z/y_p-1\\ y_l=y_{mid}/(1+\mathrm{\&eta;}l)\end{array}\right.$

In formula, y_zFor the grading brightness of image subject, X is the luminance vector of image, and W is grading luminance weights vector, y_pFor picture Evaluation brightness, η be image subject luminance errors coefficient, l be compensation coefficient, its span be 0-1 between, y_midFor brightness / 2nd of the largest component of vector, namely image intermediate luminance, y_lFor the desired level calculating, W_iFor luminance weights of grading I-th component of vector W, T represents transposition.

4. a kind of automatic explosion method judging light scene based on B-P neutral net according to claim 3, its feature It is, the evaluation brightness y of described picture_pCalculated using equation below：

y_p=X*W_z ^T

Wherein, y_pEvaluate brightness for picture, X is the luminance vector of image, W_zFor evaluating luminance weights vector.

5. a kind of automatic explosion method judging light scene based on B-P neutral net according to claim 4, its feature It is, in described step S6, described controlled amounts below equation represents：

Lnp=lntg

In formula, lnp is controlled amounts, and p=tg, t are the automatic exposure time, g analog gain coefficient.

6. a kind of automatic explosion method judging light scene based on B-P neutral net according to claim 5, its feature It is, in described step S7, time of exposure t and analog gain coefficient g is obtained using inquiry exposure gain table.