A dynamic target image intelligent analysis system and method based on deep learning

Abstract

The present invention discloses a dynamic target image intelligent analysis system and method based on deep learning, which relate to the technical field of image processing. The present invention utilizes a visual sensor and an infrared sensor to respectively collect RGB images and infrared images of a flock of birds, respectively sets weights for the RGB images and the infrared images, fuses the RGB images and the infrared images, and obtains a bionic image; generates grayscale histograms of a target area and a background area for each image block, and calculates the contrast entropy of the two grayscale histograms; obtains correction parameters for each frame of the bionic image by calculating the mean and the standard deviation; performs correction mapping on each frame of the bionic image by using the correction parameters, constructs a bionic visual filtering model, trains the bionic visual filtering model through deep learning, sets a smoothing function, smoothes each frame of the filtered bionic image, and splices all the filtered bionic image frames to obtain a final bionic image.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a system and method for intelligent analysis of dynamic target images based on deep learning. Background Art

[0002] Early bird flock monitoring relied primarily on manual observation and recording. Ornithologists used telescopes and line transects to count birds and observe their behavior within a specific area. With technological advancements, sensor-based monitoring methods, such as infrared sensors and sound sensors, have emerged. These methods indirectly obtain flock information by detecting physical signals generated by bird activity. While these methods have improved monitoring efficiency to a certain extent, they still suffer from issues such as incomplete data acquisition and low automation. With the gradual maturity of computer vision technology, some research has begun to explore its application in bird flock monitoring. Through edge detection and feature extraction in bird images, individual birds can be identified and counted. However, because traditional computer vision methods rely on handcrafted features, they are not ideal for bird flock monitoring in complex backgrounds and struggle to handle the dynamic changes of flocks, significantly limiting their application. With the continuous optimization and improvement of deep learning algorithms, such as Fast R-CNN, Faster R-CNN, YOLO, and SSD object detection algorithms, the efficiency and accuracy of bird flock monitoring have been significantly improved. At the same time, in response to the special needs of bird flock monitoring, some studies have begun to combine time series information and use RNN and its variants to analyze and predict the behavior of bird flocks.

[0003] However, when monitoring bird flocks, the birds are small in size, far away, and have different colors, and the contrast with the background color in the sky is weak, resulting in the inability of traditional single sensors to obtain clear images of the bird flock when collecting images. Summary of the Invention

[0004] The purpose of the present invention is to provide a dynamic target image intelligent analysis system and method based on deep learning to solve the problems raised in the prior art.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A method for intelligent analysis of dynamic target images based on deep learning, the method comprising the following steps:

[0007] S100, using a visual sensor and an infrared sensor to respectively collect an RGB image and an infrared image of a flock of birds, with the flock of birds in the image being used as a target area and the rest of the image except the target area being used as a background area; respectively setting weights for the RGB image and the infrared image, and fusing the RGB image and the infrared image to obtain a bionic image;

[0008] Furthermore, the specific steps for obtaining the bionic image are as follows:

[0009] S101, using visual sensors and infrared sensors to collect RGB images and infrared images of the bird flock; the spectrum data of the RGB image is obtained as P rgb , the spectral data of the infrared image is P nir ; Extract all imaging channel sets C = {R, G, B, NIR} of the two images, where R represents the RGB image P rgb The red R imaging channel in the image, G represents the RGB image P rgb The green G imaging channel in the image, B represents the RGB image P rgb The blue B imaging channel in the image, NIR represents the infrared image P nir Infrared imaging channels; calculate the significance of each imaging channel for the feather reflection of the bird flock, the formula is:

[0010] ;

[0011] In the formula, S c represents the significance of the c-th imaging channel for the feather reflection of the bird flock, Q tar Represents the total number of image pixels in the target area, P c (x, y) represents the spectral data of the intrinsic pixel (x, y) of the cth imaging channel, Q back Represents the total number of image pixels in the background area;

[0012] Calculating the significance of each imaging channel for the reflection of bird feathers can clarify the contribution of different channels (R, G, B, NIR) to the target, provide a basis for subsequent weight setting, and make the fused image more prominent in the characteristics of the bird flock.

[0013] S102. Extract the noise data in each imaging channel and calculate the noise standard deviation in each imaging channel. Normalize and standardize the significance of the bird flock feather reflection and the noise standard deviation. Use the standardized significance of the bird flock feather reflection and the noise standard deviation to calculate the weight of the corresponding imaging channel. The formula is:

[0014] ;

[0015] In the formula, w c represents the weight of the cth imaging channel, α represents the gain coefficient, and st c represents the standard deviation of noise in the cth imaging channel;

[0016] S103: Build an image fusion model using the weight of each imaging channel, specifically:

[0017] ;

[0018] In the formula, P fused represents the bionic image spectrum data obtained by fusing RGB image and infrared image, P c Represents the image spectral data of the c-th imaging channel.

[0019] Using visual sensors and infrared sensors to capture RGB and infrared images of a flock of birds, respectively, combines the color information of visible light images with the thermal information of infrared images to more comprehensively capture flock characteristics under varying lighting conditions (such as at night and in low light), improving target recognition. By weighting the RGB and infrared images, a biomimetic image is generated. Based on the saliency and noise of feather reflections in different imaging channels, image quality can be optimized, highlighting target features and enhancing image recognition and reliability.

[0020] S200, performing frame processing on the bionic image, dividing each frame of the bionic image into blocks to obtain image blocks, generating grayscale histograms of the target area and the background area for each image block, and calculating the contrast entropy of the two grayscale histograms;

[0021] Furthermore, the specific steps for calculating the contrast entropy of the two grayscale histograms are:

[0022] S201, divide the bionic image into frames, divide each frame of the bionic image into blocks to obtain image blocks, and assume that the number of image blocks is K×K. In each image block, the grayscale histogram of the target area and the background area is calculated respectively. The grayscale histogram of the target area is H tar , the grayscale histogram of the background area is H back ;

[0023] S202: Calculate the contrast entropy within each image block using the formula:

[0024] ;

[0025] In the formula, E tb represents the contrast entropy between the target area and the background area in the image block, h tar (k) represents the pixel frequency of the kth gray level in the target area within the image block, h back (k) represents the pixel frequency of the kth gray level in the background area, L represents the gray level of the image, L=255; the calculation is repeated to obtain the contrast entropy within each image block.

[0026] For each image block, grayscale histograms of the target and background areas are generated and contrast entropy is calculated to quantify the difference in grayscale distribution between the target and background. Contrast entropy can reflect the degree of distinction between the two, provide a reference for subsequent image correction, and help highlight the boundary between the target and background.

[0027] S300, calculating the mean and standard deviation of the contrast entropy of all image blocks in each frame of the bionic image, and obtaining the correction parameters of each frame of the bionic image using the mean and standard deviation;

[0028] Furthermore, the specific steps of calculating the correction parameters of each frame of bionic image using the mean and standard deviation are as follows:

[0029] S301, calculate the average value of contrast entropy of all image blocks E aver , the standard deviation is E st ; The correction parameters are calculated using the contrast entropy mean, standard deviation and contrast entropy of each image block. The formula is:

[0030] ;

[0031] In the formula, γ tb represents the correction parameter of the image block, and λ represents the hyperparameter that controls the enhancement strength;

[0032] Repeated calculations are performed to obtain correction parameters for all image blocks within a bionic image frame. Using the mean, standard deviation, and contrast entropy of each image block to calculate correction parameters, the appropriate correction strength is determined for each image block based on the statistical characteristics of the entire image frame, making the correction more targeted and adaptively enhancing the contrast between the target and the background.

[0033] S400, performing correction mapping on each frame of the bionic image using the correction parameters, constructing a bionic visual filter model, training the bionic visual filter model through deep learning, and filtering the bionic image after the correction mapping using the bionic visual filter model;

[0034] Furthermore, the specific steps of using the correction parameters to perform correction mapping on each frame of bionic image are as follows:

[0035] S401: For each frame of bionic image, each pixel in the image block is corrected using the correction parameters corresponding to the image block. The formula is:

[0036] ;

[0037] In the formula, P enh (d) represents the spectral data of the dth pixel in the image block after correction, P fused (d) represents the spectral data of the d-th pixel in the image block before correction;

[0038] S402: Correct all pixels in different image blocks using the correction parameters corresponding to the image blocks according to the formula described in S401. All corrected image blocks are then converted into a corrected bionic image. Correction mapping is performed on each bionic image frame using the correction parameters. Spectral data of the pixels can be adjusted based on the correction parameters for each image block, enhancing the difference between the target area and the background area, making the target more prominent and improving image readability.

[0039] The specific steps of filtering the bionic image after correction mapping are as follows:

[0040] S411, constructing a bionic visual filtering model, specifically:

[0041] ;

[0042] In the formula, P final (d) represents the spectral data of the d-th pixel after filtering, β represents the edge enhancement coefficient, The standard deviation is Laplace Gaussian filtering;

[0043] The corrected bionic image is filtered using a bionic visual filter model. Building a bionic visual filter model and training it through deep learning can simulate the processing methods of biological visual systems and optimize it for the characteristics of dynamic target images, making the filtering more in line with actual needs and improving the filtering effect.

[0044] S500 , setting a smoothing function, smoothing each frame of the filtered bionic image, and splicing all the filtered bionic image frames to obtain a final bionic image.

[0045] Furthermore, the specific steps of splicing all filtered bionic image frames to obtain the final bionic image are as follows:

[0046] S501, set the spectrum change threshold Pτ, and calculate the spectrum data difference between adjacent frames after bionic image filtering in real time. The formula is:

[0047] ;

[0048] In the formula, D(t) represents the difference in spectral data after bionic image filtering between the tth frame and the t-1th frame, u c （t） Indicates the mean spectral data of all pixels in the target area in the cth imaging channel at the tth frame, u c （t-1） represents the mean spectral data of all pixels in the target area in the cth imaging channel at the t-1th frame;

[0049] S502, using a spectral change threshold to judge the calculated spectral data difference of adjacent frame images, when D(t) ≥ Pτ, the filtered bionic image is judged to be non-smooth, and a smoothing mechanism is activated; when the spectral data differences of all adjacent frame images satisfy D(t) < Pτ, the filtered bionic image is judged to be smooth;

[0050] S503, set the smoothing mechanism, , in the formula, P final _h represents the final bionic image after smoothing, a represents the smoothing factor, P final t represents the spectral data of the t-th frame bionic image, P final t-1 Represents the spectral data of the t-1th frame of the bionic image.

[0051] When the spectral differences between adjacent frames are large, the smoothing mechanism is activated. The filtered frames are smoothed by the smoothing function, which can reduce the mutations between frames, make the image sequence transition more natural, avoid image flickering or discontinuity caused by dynamic changes, and improve the stability of dynamic target analysis.

[0052] The final image is obtained by stitching all the smoothed bionic image frames together, which can integrate the information of multiple frames to form a more complete and clear dynamic target image, making it easier to analyze the overall characteristics and motion trajectory of the target.

[0053] A dynamic target image intelligent analysis system based on deep learning, the dynamic target image intelligent analysis system includes an image acquisition module, an image fusion module, an image segmentation module, an image comparison module, an image processing module and an image optimization module;

[0054] The image acquisition module is used to collect RGB images and infrared images of the bird flock using a visual sensor and an infrared sensor respectively;

[0055] The image fusion module is used to set the weights of the RGB image and the infrared image respectively, and fuse the RGB image and the infrared image to obtain a bionic image;

[0056] The image segmentation module is used to perform frame processing on the bionic image, and to perform block processing on each frame of the bionic image to obtain image blocks;

[0057] The image contrast module is used to generate grayscale histograms of the target area and the background area for each image block, and calculate the contrast entropy of the two grayscale histograms;

[0058] The image processing module is used to calculate the mean and standard deviation of the contrast entropy of all image blocks in each frame of bionic image, and use the mean and standard deviation to calculate the correction parameters of each frame of bionic image; and use the correction parameters to perform correction mapping on each frame of bionic image;

[0059] The image optimization module is used to construct a bionic visual filtering model, train the bionic visual filtering model through deep learning, and use the bionic visual filtering model to filter the bionic image after correction mapping; set a smoothing function to smooth each frame of the filtered bionic image, and splice all the filtered bionic image frames to obtain the final bionic image.

[0060] The image fusion module includes a weighted fusion unit and a weight calculation unit;

[0061] The weighted fusion unit is used to perform weighted fusion of images of all imaging channels using the weight of each imaging channel;

[0062] The weight calculation unit is used to calculate the significance of each imaging channel for the feather reflection of the bird flock, and calculate the weight of the corresponding imaging channel using the significance of the feather reflection of the bird flock and the noise standard deviation.

[0063] The image optimization module includes a filtering unit and a smoothing unit;

[0064] The filtering unit is used to construct a bionic visual filtering model, and use the bionic visual filtering model to filter the corrected bionic image;

[0065] The smoothing unit is used to set a smoothing function, perform smoothing on each frame of filtered bionic image, and splice all filtered bionic image frames to obtain a final bionic image.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] 1. This invention combines the advantages of RGB images and infrared images. Through weight setting and fusion, the generated bionic images have richer information and can better present target features under different lighting conditions, thereby improving the quality and reliability of the image.

[0068] 2. From framing and block division to contrast entropy calculation and then to correction mapping determination, the present invention performs adaptive processing based on the local and overall features of the image, effectively enhancing the contrast between the target and the background, making the target more prominent and facilitating subsequent recognition and analysis.

[0069] 3. By setting the spectral change threshold to judge the difference between adjacent frames and starting the smoothing mechanism for processing, the smooth transition of the image sequence is ensured, the image discontinuity problem caused by dynamic changes is avoided, and the system can more stably analyze the motion characteristics and trajectory of dynamic targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1This is a module distribution diagram of a dynamic target image intelligent analysis system based on deep learning in the present invention;

[0071] Figure 2 This is a schematic diagram of the steps of a deep learning-based intelligent analysis method for dynamic target images in the present invention. DETAILED DESCRIPTION

[0072] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0073] Example: Figure 1-Figure 2 As shown, the present invention provides a technical solution.

[0074] A method for intelligent analysis of dynamic target images based on deep learning, the method comprising the following steps:

[0075] S100, using a visual sensor and an infrared sensor to respectively collect an RGB image and an infrared image of a flock of birds, with the flock of birds in the image being used as a target area and the rest of the image except the target area being used as a background area; respectively setting weights for the RGB image and the infrared image, and fusing the RGB image and the infrared image to obtain a bionic image;

[0076] The specific steps to obtain the bionic image are:

[0077] S101, using visual sensors and infrared sensors to collect RGB images and infrared images of the bird flock; the spectrum data of the RGB image is obtained as P rgb , the spectral data of the infrared image is P nir ; Extract all imaging channel sets C = {R, G, B, NIR} of the two images, where R represents the RGB image P rgb The red R imaging channel in the image, G represents the RGB image P rgb The green G imaging channel in the image, B represents the RGB image P rgb The blue B imaging channel in the image, NIR represents the infrared image P nir Infrared imaging channels; calculate the significance of each imaging channel for the feather reflection of the bird flock, the formula is:

[0078] ;

[0079] In the formula, S c represents the significance of the c-th imaging channel for the feather reflection of the bird flock, Q tarRepresents the total number of image pixels in the target area, P c (x, y) represents the spectral data of the intrinsic pixel (x, y) of the cth imaging channel, Q back Represents the total number of image pixels in the background area;

[0080] Calculating the significance of each imaging channel for the reflection of bird feathers can clarify the contribution of different channels (R, G, B, NIR) to the target, provide a basis for subsequent weight setting, and make the fused image more prominent in the characteristics of the bird flock.

[0081] S102. Extract the noise data in each imaging channel and calculate the noise standard deviation in each imaging channel. Normalize and standardize the significance of the bird flock feather reflection and the noise standard deviation. Use the standardized significance of the bird flock feather reflection and the noise standard deviation to calculate the weight of the corresponding imaging channel. The formula is:

[0082] ;

[0083] In the formula, w c represents the weight of the cth imaging channel, α represents the gain coefficient, and st c represents the standard deviation of noise in the cth imaging channel;

[0084] S103: Build an image fusion model using the weight of each imaging channel, specifically:

[0085] ;

[0086] In the formula, P fused represents the bionic image spectrum data obtained by fusing RGB image and infrared image, P c Represents the image spectral data of the c-th imaging channel.

[0087] Using visual sensors and infrared sensors to capture RGB and infrared images of a flock of birds, respectively, combines the color information of visible light images with the thermal information of infrared images to more comprehensively capture flock characteristics under varying lighting conditions (such as at night and in low light), improving target recognition. By weighting the RGB and infrared images, a biomimetic image is generated. Based on the saliency and noise of feather reflections in different imaging channels, image quality can be optimized, highlighting target features and enhancing image recognition and reliability.

[0088] S200, performing frame processing on the bionic image, dividing each frame of the bionic image into blocks to obtain image blocks, generating grayscale histograms of the target area and the background area for each image block, and calculating the contrast entropy of the two grayscale histograms;

[0089] The specific steps for calculating the contrast entropy of two grayscale histograms are:

[0090] S201, divide the bionic image into frames, divide each frame of the bionic image into blocks to obtain image blocks, and assume that the number of image blocks is K×K. In each image block, the grayscale histogram of the target area and the background area is calculated respectively. The grayscale histogram of the target area is H tar , the grayscale histogram of the background area is H back ;

[0091] S202: Calculate the contrast entropy within each image block using the formula:

[0092] ;

[0093] In the formula, E tb represents the contrast entropy between the target area and the background area in the image block, h tar (k) represents the pixel frequency of the kth gray level in the target area within the image block, h back (k) represents the pixel frequency of the kth gray level in the background area, L represents the gray level of the image, L=255; the calculation is repeated to obtain the contrast entropy within each image block.

[0094] For each image block, grayscale histograms of the target and background areas are generated and contrast entropy is calculated to quantify the difference in grayscale distribution between the target and background. Contrast entropy can reflect the degree of distinction between the two, provide a reference for subsequent image correction, and help highlight the boundary between the target and background.

[0095] S300, calculating the mean and standard deviation of the contrast entropy of all image blocks in each frame of the bionic image, and obtaining the correction parameters of each frame of the bionic image using the mean and standard deviation;

[0096] The specific steps for calculating the correction parameters of each frame of bionic image using the mean and standard deviation are as follows:

[0097] S301, calculate the average value of contrast entropy of all image blocks E aver , the standard deviation is E st ; The correction parameters are calculated using the contrast entropy mean, standard deviation and contrast entropy of each image block. The formula is:

[0098] ;

[0099] In the formula, γ tb represents the correction parameter of the image block, and λ represents the hyperparameter that controls the enhancement strength;

[0100] Repeated calculations are performed to obtain correction parameters for all image blocks within a bionic image frame. Using the mean, standard deviation, and contrast entropy of each image block to calculate correction parameters, the appropriate correction strength is determined for each image block based on the statistical characteristics of the entire image frame, making the correction more targeted and adaptively enhancing the contrast between the target and the background.

[0101] S400, performing correction mapping on each frame of the bionic image using the correction parameters, constructing a bionic visual filter model, training the bionic visual filter model through deep learning, and filtering the bionic image after the correction mapping using the bionic visual filter model;

[0102] The specific steps of using the correction parameters to perform correction mapping on each frame of bionic image are as follows:

[0103] S401: For each frame of bionic image, each pixel in the image block is corrected using the correction parameters corresponding to the image block. The formula is:

[0104] ;

[0105] In the formula, P enh (d) represents the spectral data of the dth pixel in the image block after correction, P fused (d) represents the spectral data of the d-th pixel in the image block before correction;

[0106] S402: Correct all pixels in different image blocks using the correction parameters corresponding to the image blocks according to the formula described in S401. All corrected image blocks are then converted into a corrected bionic image. Correction mapping is performed on each bionic image frame using the correction parameters. Spectral data of the pixels can be adjusted based on the correction parameters for each image block, enhancing the difference between the target area and the background area, making the target more prominent and improving image readability.

[0107] The specific steps of filtering the bionic image after correction mapping are as follows:

[0108] S411, constructing a bionic visual filtering model, specifically:

[0109] ;

[0110] In the formula, P final (d) represents the spectral data of the d-th pixel after filtering, β represents the edge enhancement coefficient, The standard deviation is Laplace Gaussian filtering;

[0111] The corrected bionic image is filtered using a bionic visual filter model. Building a bionic visual filter model and training it through deep learning can simulate the processing methods of biological visual systems and optimize it for the characteristics of dynamic target images, making the filtering more in line with actual needs and improving the filtering effect.

[0112] S500 , setting a smoothing function, smoothing each frame of the filtered bionic image, and splicing all the filtered bionic image frames to obtain a final bionic image.

[0113] The specific steps of stitching all filtered bionic image frames to obtain the final bionic image are as follows:

[0114] S501, set the spectrum change threshold Pτ, and calculate the spectrum data difference between adjacent frames after bionic image filtering in real time. The formula is:

[0115] ;

[0116] In the formula, D(t) represents the difference in spectral data after bionic image filtering between the tth frame and the t-1th frame, u c （t） Indicates the mean spectral data of all pixels in the target area in the cth imaging channel at the tth frame, u c （t-1） represents the mean spectral data of all pixels in the target area in the cth imaging channel at the t-1th frame;

[0117] S502, using a spectral change threshold to judge the calculated spectral data difference of adjacent frame images, when D(t) ≥ Pτ, the filtered bionic image is judged to be non-smooth, and a smoothing mechanism is activated; when the spectral data differences of all adjacent frame images satisfy D(t) < Pτ, the filtered bionic image is judged to be smooth;

[0118] S503, set the smoothing mechanism, , in the formula, P final _h represents the final bionic image after smoothing, a represents the smoothing factor, P final t represents the spectral data of the t-th frame bionic image, P final t-1 Represents the spectral data of the t-1th frame of the bionic image.

[0119] When the spectral differences between adjacent frames are large, the smoothing mechanism is activated. The filtered frames are smoothed by the smoothing function, which can reduce the mutations between frames, make the image sequence transition more natural, avoid image flickering or discontinuity caused by dynamic changes, and improve the stability of dynamic target analysis.

[0120] The final image is obtained by stitching all the smoothed bionic image frames together, which can integrate the information of multiple frames to form a more complete and clear dynamic target image, making it easier to analyze the overall characteristics and motion trajectory of the target.

[0121] A dynamic target image intelligent analysis system based on deep learning, the dynamic target image intelligent analysis system includes an image acquisition module, an image fusion module, an image segmentation module, an image comparison module, an image processing module and an image optimization module;

[0122] The image acquisition module is used to collect RGB images and infrared images of the bird flock using a visual sensor and an infrared sensor respectively;

[0123] The image fusion module is used to set the weights of the RGB image and the infrared image respectively, and fuse the RGB image and the infrared image to obtain a bionic image;

[0124] The image segmentation module is used to perform frame processing on the bionic image, and to perform block processing on each frame of the bionic image to obtain image blocks;

[0125] The image contrast module is used to generate grayscale histograms of the target area and the background area for each image block, and calculate the contrast entropy of the two grayscale histograms;

[0126] The image processing module is used to calculate the mean and standard deviation of the contrast entropy of all image blocks in each frame of bionic image, and use the mean and standard deviation to calculate the correction parameters of each frame of bionic image; and use the correction parameters to perform correction mapping on each frame of bionic image;

[0127] The image optimization module is used to construct a bionic visual filtering model, train the bionic visual filtering model through deep learning, and use the bionic visual filtering model to filter the bionic image after correction mapping; set a smoothing function to smooth each frame of the filtered bionic image, and splice all the filtered bionic image frames to obtain the final bionic image.

[0128] The image fusion module includes a weighted fusion unit and a weight calculation unit;

[0129] The weighted fusion unit is used to perform weighted fusion of images of all imaging channels using the weight of each imaging channel;

[0130] The weight calculation unit is used to calculate the significance of each imaging channel for the feather reflection of the bird flock, and calculate the weight of the corresponding imaging channel using the significance of the feather reflection of the bird flock and the noise standard deviation.

[0131] The image optimization module includes a filtering unit and a smoothing unit;

[0132] The filtering unit is used to construct a bionic visual filtering model, and use the bionic visual filtering model to filter the corrected bionic image;

[0133] The smoothing unit is used to set a smoothing function, perform smoothing on each frame of filtered bionic image, and splice all filtered bionic image frames to obtain a final bionic image.

[0134] Example: For real-time dynamic monitoring of bird flocks at an airport, infrared sensors and visual sensors are set up. In the real-time scene, there is a group of pigeons flying from the side of the runway to the takeoff path. The significance of the bird flock feather reflection in each imaging channel is calculated separately. It is found that the pigeon feathers have stronger reflection in the blue channel. B =25, infrared channel high reflection is S NIR =40;

[0135] Calculate the weights for image weighted fusion, and calculate the blue channel weight to be 0.29 and the infrared channel weight to be 0.71. Use the weights for fusion to obtain the bionic image.

[0136] The bionic image is processed by dividing frames into blocks. The contrast entropy in one image block is calculated to be 3.5. The average contrast entropy is set to 2.3 and the standard deviation is set to 0.8. The correction parameter in the corresponding image block is calculated to be 0.45.

[0137] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A method for intelligent analysis of dynamic target images based on deep learning, characterized by: The method comprises the following steps: S100, using a visual sensor and an infrared sensor to respectively collect an RGB image and an infrared image of a flock of birds, with the flock of birds in the image being used as a target area and the rest of the image except the target area being used as a background area; respectively setting weights for the RGB image and the infrared image, and fusing the RGB image and the infrared image to obtain a bionic image; S200, performing frame processing on the bionic image, dividing each frame of the bionic image into blocks to obtain image blocks, generating grayscale histograms of the target area and the background area for each image block, and calculating the contrast entropy of the two grayscale histograms; The specific steps for calculating the contrast entropy of two grayscale histograms are: S201, divide the bionic image into frames, divide each frame of the bionic image into blocks to obtain image blocks, and assume that the number of image blocks is K×K. In each image block, the grayscale histogram of the target area and the background area is calculated respectively. The grayscale histogram of the target area is H tar , the grayscale histogram of the background area is H back ; S202: Calculate the contrast entropy within each image block using the formula: ； In the formula, E tb represents the contrast entropy between the target area and the background area in the image block, h tar (k) represents the pixel frequency of the kth gray level in the target area within the image block, h back (k) represents the pixel frequency of the kth gray level in the background area, L represents the gray level of the image, L=255; repeat the calculation to obtain the contrast entropy of each image block; S300, calculating the mean and standard deviation of the contrast entropy of all image blocks in each frame of the bionic image, and obtaining the correction parameters of each frame of the bionic image using the mean and standard deviation; S400, performing correction mapping on each frame of the bionic image using the correction parameters, constructing a bionic visual filter model, training the bionic visual filter model through deep learning, and filtering the bionic image after the correction mapping using the bionic visual filter model; S500 , setting a smoothing function, smoothing each frame of the filtered bionic image, and splicing all the filtered bionic image frames to obtain a final bionic image.

2. The method for intelligent analysis of dynamic target images based on deep learning according to claim 1, characterized in that: The specific steps of obtaining the bionic image in S100 are: S101, using visual sensors and infrared sensors to collect RGB images and infrared images of the bird flock; the spectrum data of the RGB image is obtained as P rgb , the spectral data of the infrared image is P nir ; Extract all imaging channel sets C = {R, G, B, NIR} of the two images, where R represents the RGB image P rgb The red R imaging channel in the image, G represents the RGB image P rgb The green G imaging channel in the image, B represents the RGB image P rgb The blue B imaging channel in the image, NIR represents the infrared image P nir Infrared imaging channels; calculate the significance of each imaging channel for the feather reflection of the bird flock, the formula is: ； In the formula, S c represents the significance of the c-th imaging channel for the feather reflection of the bird flock, Q tar Represents the total number of image pixels in the target area, P c (x, y) represents the spectral data of the intrinsic pixel (x, y) of the cth imaging channel, Q back Represents the total number of image pixels in the background area; S102. Extract the noise data in each imaging channel and calculate the noise standard deviation in each imaging channel. Normalize and standardize the significance of the bird flock feather reflection and the noise standard deviation. Use the standardized significance of the bird flock feather reflection and the noise standard deviation to calculate the weight of the corresponding imaging channel. The formula is: ； In the formula, w c represents the weight of the cth imaging channel, α represents the gain coefficient, and st c represents the standard deviation of noise in the cth imaging channel; S103: Build an image fusion model using the weight of each imaging channel, specifically: ； In the formula, P fused represents the bionic image spectrum data obtained by fusing RGB image and infrared image, P c Represents the image spectral data of the c-th imaging channel.

3. The method for intelligent analysis of dynamic target images based on deep learning according to claim 2, characterized in that: The specific steps of calculating the correction parameters of each frame of bionic image using the mean and standard deviation in S300 are: S301, calculate the average value of contrast entropy of all image blocks E aver , the standard deviation is E st ; The correction parameters are calculated using the contrast entropy mean, standard deviation and contrast entropy of each image block. The formula is: ； In the formula, γ tb represents the correction parameter of the image block, and λ represents the hyperparameter that controls the enhancement strength; Repeat the calculation to obtain the correction parameters of all image blocks in a frame of bionic image.

4. The method for intelligent analysis of dynamic target images based on deep learning according to claim 3, characterized in that: The specific steps of performing correction mapping on each frame of bionic image using correction parameters in S400 are: S401: For each frame of bionic image, each pixel in the image block is corrected using the correction parameters corresponding to the image block. The formula is: ； In the formula, P enh (d) represents the spectral data of the dth pixel in the image block after correction, P fused (d) represents the spectral data of the d-th pixel in the image block before correction; S402 , correcting all pixels in different image blocks using correction parameters corresponding to the image blocks according to the formula in S401 , and obtaining all corrected image blocks to form a frame of corrected bionic image.

5. The method for intelligent analysis of dynamic target images based on deep learning according to claim 4, characterized in that: The specific steps of filtering the corrected and mapped bionic image in S400 are: S411, constructing a bionic visual filtering model, specifically: ； In the formula, P final (d) represents the spectral data of the d-th pixel after filtering, β represents the edge enhancement coefficient, The standard deviation is Laplace Gaussian filtering; The corrected bionic image is filtered using a bionic visual filtering model.

6. The method for intelligent analysis of dynamic target images based on deep learning according to claim 5, characterized in that: The specific steps of stitching all filtered bionic image frames to obtain the final bionic image in S500 are: S501, set the spectrum change threshold Pτ, and calculate the spectrum data difference between adjacent frames after bionic image filtering in real time. The formula is: ； In the formula, D(t) represents the difference in spectral data after bionic image filtering between the tth frame and the t-1th frame, u c （t） Indicates the mean spectral data of all pixels in the target area in the cth imaging channel at the tth frame, u c （t-1） represents the mean spectral data of all pixels in the target area in the cth imaging channel at the t-1th frame; S502, using a spectral change threshold to judge the calculated spectral data difference of adjacent frame images, when D(t) ≥ Pτ, the filtered bionic image is judged to be non-smooth, and a smoothing mechanism is activated; when the spectral data differences of all adjacent frame images satisfy D(t) < Pτ, the filtered bionic image is judged to be smooth; S503, set the smoothing mechanism, , in the formula, P final _h represents the final bionic image after smoothing, a represents the smoothing factor, P final t represents the spectral data of the t-th frame bionic image, P final t-1 Represents the spectral data of the t-1th frame of the bionic image.

7. A system for intelligent analysis of dynamic target images based on deep learning, applying the method for intelligent analysis of dynamic target images based on deep learning according to any one of claims 1 to 6, characterized in that: The dynamic target image intelligent analysis system includes image acquisition module, image fusion module, image segmentation module, image comparison module, image processing module and image optimization module; The image acquisition module is used to collect RGB images and infrared images of the bird flock using a visual sensor and an infrared sensor respectively; The image fusion module is used to set the weights of the RGB image and the infrared image respectively, and fuse the RGB image and the infrared image to obtain a bionic image; The image segmentation module is used to perform frame processing on the bionic image, and to perform block processing on each frame of the bionic image to obtain image blocks; The image contrast module is used to generate grayscale histograms of the target area and the background area for each image block, and calculate the contrast entropy of the two grayscale histograms; The image processing module is used to calculate the mean and standard deviation of the contrast entropy of all image blocks in each frame of bionic image, and use the mean and standard deviation to calculate the correction parameters of each frame of bionic image; and use the correction parameters to perform correction mapping on each frame of bionic image; The image optimization module is used to construct a bionic visual filtering model, train the bionic visual filtering model through deep learning, and use the bionic visual filtering model to filter the bionic image after correction mapping; set a smoothing function to smooth each frame of the filtered bionic image, and splice all the filtered bionic image frames to obtain the final bionic image.

8. The deep learning-based dynamic target image intelligent analysis system according to claim 7, characterized in that: The image fusion module includes a weighted fusion unit and a weight calculation unit; The weighted fusion unit is used to perform weighted fusion of images of all imaging channels using the weight of each imaging channel; The weight calculation unit is used to calculate the significance of each imaging channel for the feather reflection of the bird flock, and calculate the weight of the corresponding imaging channel using the significance of the feather reflection of the bird flock and the noise standard deviation.

9. The deep learning-based dynamic target image intelligent analysis system according to claim 7, characterized in that: The image optimization module includes a filtering unit and a smoothing unit; The filtering unit is used to construct a bionic visual filtering model, and use the bionic visual filtering model to filter the corrected bionic image; The smoothing unit is used to set a smoothing function, perform smoothing on each frame of filtered bionic image, and splice all filtered bionic image frames to obtain a final bionic image.

Images