An intelligent algorithm and model based on AI chip to support intelligent bird identification

Abstract

The present invention discloses an intelligent algorithm and model based on AI chip to support intelligent bird recognition, including: collecting an initial image including birds; standardizing the size of the initial image to obtain a target image and inputting it into a deep convolutional neural network for convolution operation to obtain image features; matching the image features with preset image features in a database, and determining category information according to the matching results. Intelligent bird recognition is achieved and recognition accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an intelligent algorithm and model based on an AI chip that supports intelligent bird identification. Background Art

[0002] With the continuous development of Internet of Things technology, relevant images can be obtained and image recognition can be performed, and corresponding measures can be determined based on the recognition results. In the existing technology, intelligent recognition of birds cannot be achieved, and the recognition efficiency and accuracy are not high. Summary of the invention

[0003] The present invention aims to solve at least one of the technical problems in the above-mentioned technology to a certain extent. To this end, the first purpose of the present invention is to propose an intelligent algorithm based on an AI chip to support intelligent identification of birds, which facilitates the intelligent identification of birds and improves the identification efficiency and accuracy.

[0004] The second object of the present invention is to propose an intelligent model based on AI chip to support intelligent identification of birds.

[0005] To achieve the above-mentioned purpose, the first embodiment of the present invention proposes an intelligent algorithm based on an AI chip to support intelligent identification of birds, including:

[0006] The acquisition includes initial images of the birds;

[0007] The size of the initial image is normalized to obtain a target image and the target image is input into a deep convolutional neural network for convolution operation to obtain image features;

[0008] The image features are matched with preset image features in a database, and category information is determined according to the matching results.

[0009] According to some embodiments of the present invention, before inputting the target image into the deep convolutional neural network for convolution operation, the method further includes:

[0010] Performing grayscale processing on the target image to obtain a grayscale image;

[0011] Obtaining the grayscale values ​​of the pixels in the grayscale image, screening out the maximum grayscale value and the minimum grayscale value, and calculating the grayscale mean and standard deviation;

[0012] Calculating the addition value and subtraction value of the grayscale mean and standard deviation;

[0013] Compare the maximum grayscale value with the added value, and determine the larger value as the first calculated value;

[0014] Compare the minimum grayscale value with the subtraction value, and determine the smaller value as the second calculated value;

[0015] Calculating a ratio of the second calculated value to the first calculated value, and determining whether the ratio is less than a preset ratio;

[0016] When it is determined that the ratio is less than a preset ratio, Gaussian blur processing is performed on the grayscale image to determine the grayscale value of each pixel in the grayscale image after the Gaussian blur processing;

[0017] Determine the grayscale difference of each pixel in the grayscale image before and after Gaussian blur processing, determine the adjustment parameter of the corresponding pixel according to the grayscale difference, and perform adjustment processing on the corresponding pixel according to the adjustment parameter.

[0018] According to some embodiments of the present invention, before inputting the target image into the deep convolutional neural network for convolution operation, the method further includes:

[0019] Scoring the target image based on a number of image quality evaluation indicators to obtain a corresponding single score; calculating according to the single score and preset evaluation weights of the number of image quality evaluation indicators to obtain a comprehensive score, and determining whether the comprehensive score is less than the preset score;

[0020] When it is determined that the comprehensive score is less than the preset score, the target image is analyzed to determine a local image including texture features and smooth features;

[0021] Matching the local image with a preset local image in a preset image library, and determining a correction parameter according to the matching result;

[0022] Performing noise reduction and smoothing processing on the local image according to the correction parameters to obtain a corrected image;

[0023] Processing the corrected image based on an unsharp mask algorithm to obtain an edge enhanced image;

[0024] Performing edge extraction processing on the edge enhanced image based on the Canny operator and generating a binary image;

[0025] Determine edge direction information of each edge pixel point according to the binary image, and determine a first edge distribution map according to the edge direction information;

[0026] Setting a sliding window on the first edge distribution map, determining a central pixel point of the sliding image included in the sliding window, performing a difference operation between the coordinates of other pixel points in the sliding image except the central pixel point and the coordinates of the central pixel point to obtain a centralized coordinate;

[0027] According to the centralized coordinates and the edge direction information of the central pixel, an association relationship between other pixel points and the central pixel point is established; and a second edge distribution map is determined according to the association relationship.

[0028] According to some embodiments of the present invention, the deep convolutional neural network includes 2 convolutional layers and a computing layer, 2 fully connected layers, 1 input layer and 1 output layer; the sigmoid function is used as the activation function in the convolutional layer; the computing layer is connected to the convolutional layer to calculate the local average and feature extraction; the fully connected layer is used to connect the features determined by the computing layer; the output layer is used to output the final image features.

[0029] According to some embodiments of the present invention, the image features include contour features.

[0030] According to some embodiments of the present invention, matching the image features with preset image features in a database, and determining category information according to the matching results, includes:

[0031] The matching degrees of the image features and the preset image features in the database are calculated respectively, the preset image feature with the greatest matching degree is determined, and the bird information corresponding to the preset image feature with the greatest matching degree is determined as the category information corresponding to the initial image.

[0032] According to some embodiments of the present invention, before inputting the target image into the deep convolutional neural network for convolution operation, the method further includes:

[0033] Input the target image into a U-Net model, perform connected domain analysis on the target image, and obtain a plurality of connected domain images;

[0034] Acquire the areas of several connected domain images, and compare them with preset areas respectively, and use the connected domain images with areas larger than the preset areas as the images to be processed;

[0035] Determine the output probability of all pixels on the image to be processed based on the U-Net model and perform adaptive processing based on preset rules to obtain the corrected probability and compare it with the preset probability;

[0036] The pixel points whose correction probability is greater than the preset probability are taken as reserved pixel points, and a reserved image is obtained according to the reserved pixel points;

[0037] Performing histogram equalization processing on the retained image to obtain an enhanced image;

[0038] Performing threshold processing on the enhanced image based on the maximum inter-class variance method to obtain a grayscale segmentation threshold for image segmentation of the enhanced image;

[0039] Performing binarization processing according to the grayscale segmentation threshold and the grayscale value of each pixel in the enhanced image, and determining the posture of the bird according to the binarization processing result;

[0040] The enhanced image marked with the bird's posture is input into the deep convolutional neural network for convolution operation.

[0041] To achieve the above-mentioned purpose, the second embodiment of the present invention proposes an intelligent model based on an AI chip to support intelligent identification of birds, including:

[0042] A collection module, used for collecting initial images including birds;

[0043] A processing module, used for normalizing the size of the initial image to obtain a target image and inputting the target image into a deep convolutional neural network for convolution operation to obtain image features;

[0044] The matching module is used to match the image features with preset image features in the database and determine the category information according to the matching results.

[0045] According to some embodiments of the present invention, the further comprising:

[0046] Regulation modules for:

[0047] Before the processing module inputs the target image into the deep convolutional neural network for convolution operation, grayscale processing is performed on the target image to obtain a grayscale image;

[0048] Obtaining the grayscale values ​​of the pixels in the grayscale image, screening out the maximum grayscale value and the minimum grayscale value, and calculating the grayscale mean and standard deviation;

[0049] Calculating the addition value and subtraction value of the grayscale mean and standard deviation;

[0050] Compare the maximum grayscale value with the added value, and determine the larger value as the first calculated value;

[0051] Compare the minimum grayscale value with the subtraction value, and determine the smaller value as the second calculated value;

[0052] Calculating a ratio of the second calculated value to the first calculated value, and determining whether the ratio is less than a preset ratio;

[0053] When it is determined that the ratio is less than a preset ratio, Gaussian blur processing is performed on the grayscale image to determine the grayscale value of each pixel in the grayscale image after the Gaussian blur processing;

[0054] Determine the grayscale difference of each pixel in the grayscale image before and after Gaussian blur processing, determine the adjustment parameter of the corresponding pixel according to the grayscale difference, and perform adjustment processing on the corresponding pixel according to the adjustment parameter.

[0055] According to some embodiments of the present invention, the further comprising:

[0056] The correction module is used to score the target image based on several image quality evaluation indicators before the processing module inputs the target image into the deep convolutional neural network for convolution operation, and obtain the corresponding single score; calculate according to the single score and the preset evaluation weights of several image quality evaluation indicators to obtain a comprehensive score, and determine whether it is less than the preset score;

[0057] When it is determined that the comprehensive score is less than the preset score, the target image is analyzed to determine a local image including texture features and smooth features;

[0058] Matching the local image with a preset local image in a preset image library, and determining a correction parameter according to the matching result;

[0059] Performing noise reduction and smoothing processing on the local image according to the correction parameters to obtain a corrected image;

[0060] Processing the corrected image based on an unsharp mask algorithm to obtain an edge enhanced image;

[0061] Performing edge extraction processing on the edge enhanced image based on the Canny operator and generating a binary image;

[0062] Determine edge direction information of each edge pixel point according to the binary image, and determine a first edge distribution map according to the edge direction information;

[0063] Setting a sliding window on the first edge distribution map, determining a central pixel point of the sliding image included in the sliding window, performing a difference operation between the coordinates of other pixel points in the sliding image except the central pixel point and the coordinates of the central pixel point to obtain a centralized coordinate;

[0064] According to the centralized coordinates and the edge direction information of the central pixel, an association relationship between other pixel points and the central pixel point is established; and a second edge distribution map is determined according to the association relationship.

[0065] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0066] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0068] Figure 1 is a flow chart of an intelligent algorithm supporting intelligent bird identification based on an AI chip according to an embodiment of the present invention;

[0069] Figure 2 is a flow chart of an intelligent algorithm supporting intelligent bird identification based on an AI chip according to yet another embodiment of the present invention;

[0070] Figure 3 It is a block diagram of an intelligent model supporting intelligent bird identification based on an AI chip according to an embodiment of the present invention. DETAILED DESCRIPTION

[0071] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0072] like Figure 1 As shown, the first embodiment of the present invention proposes an intelligent algorithm based on an AI chip to support intelligent identification of birds, including steps S1-S3:

[0073] S1. Collecting initial images including birds;

[0074] S2, normalizing the size of the initial image to obtain a target image and inputting the target image into a deep convolutional neural network for convolution operation to obtain image features;

[0075] S3. Match the image features with preset image features in a database, and determine category information according to the matching results.

[0076] The working principle of the above technical solution is as follows: collecting an initial image including birds; standardizing the size of the initial image to obtain a target image and inputting it into a deep convolutional neural network for convolution operation to obtain image features; matching the image features with preset image features in a database, and determining category information based on the matching results.

[0077] The beneficial effects of the above technical solution are as follows: the size of the initial image is standardized to meet the input parameter requirements of the deep convolutional neural network for convolution operations, which facilitates obtaining accurate image features, and then determines the category information based on the matching results, thereby facilitating the intelligent recognition of birds and improving recognition efficiency and accuracy.

[0078] like Figure 2 As shown, according to some embodiments of the present invention, before the target image is input into the deep convolutional neural network for convolution operation, steps S21-S28 are also included:

[0079] S21, performing grayscale processing on the target image to obtain a grayscale image;

[0080] S22, obtaining the grayscale values ​​of the pixels in the grayscale image, screening out the maximum grayscale value and the minimum grayscale value, and calculating the grayscale mean and standard deviation;

[0081] S23, calculating the addition value and subtraction value of the grayscale mean and standard deviation;

[0082] S24, comparing the maximum gray value with the added value, and determining the larger value as the first calculated value;

[0083] S25, comparing the minimum grayscale value with the subtraction value, and determining the smaller value as the second calculated value;

[0084] S26, calculating the ratio of the second calculated value to the first calculated value, and determining whether the ratio is less than a preset ratio;

[0085] S27, when it is determined that the ratio is less than the preset ratio, performing Gaussian blur processing on the grayscale image, and determining the grayscale value of each pixel in the grayscale image after the Gaussian blur processing;

[0086] S28, determining the grayscale difference of each pixel in the grayscale image before and after the Gaussian blur processing, determining the adjustment parameter of the corresponding pixel according to the grayscale difference, and performing adjustment processing on the corresponding pixel according to the adjustment parameter.

[0087] The working principle of the above technical solution is as follows: grayscale processing is performed on the target image to obtain a grayscale image; the grayscale values ​​of the pixels in the grayscale image are obtained, the maximum grayscale value and the minimum grayscale value are screened out, and the grayscale mean and standard deviation are calculated; the addition value and subtraction value of the grayscale mean and standard deviation are calculated; the maximum grayscale value is compared with the addition value, and the larger value is determined as the first calculation value; the minimum grayscale value is compared with the subtraction value, and the smaller value is determined as the second calculation value; the ratio of the second calculation value to the first calculation value is calculated, and it is determined whether it is less than the preset ratio; the ratio is characterized as clarity. When it is determined that the ratio is less than the preset ratio, Gaussian blur processing is performed on the grayscale image to determine the grayscale value of each pixel in the grayscale image after Gaussian blur processing; the grayscale difference of each pixel in the grayscale image before and after Gaussian blur processing is determined, and the adjustment parameters of the corresponding pixel are determined according to the grayscale difference, and the corresponding pixel is adjusted according to the adjustment parameters.

[0088] The beneficial effect of the above technical solution is: when it is determined that the clarity of the grayscale image is low, clarity adjustment processing is performed to ensure the clarity of the input image, thereby facilitating improving the recognition accuracy based on the deep convolutional neural network.

[0089] According to some embodiments of the present invention, before inputting the target image into the deep convolutional neural network for convolution operation, the method further includes:

[0090] Scoring the target image based on a number of image quality evaluation indicators to obtain a corresponding single score; calculating according to the single score and preset evaluation weights of the number of image quality evaluation indicators to obtain a comprehensive score, and determining whether the comprehensive score is less than the preset score;

[0091] When it is determined that the comprehensive score is less than the preset score, the target image is analyzed to determine a local image including texture features and smooth features;

[0092] Matching the local image with a preset local image in a preset image library, and determining a correction parameter according to the matching result;

[0093] Performing noise reduction and smoothing processing on the local image according to the correction parameters to obtain a corrected image;

[0094] Processing the corrected image based on an unsharp mask algorithm to obtain an edge enhanced image;

[0095] Performing edge extraction processing on the edge enhanced image based on the Canny operator and generating a binary image;

[0096] Determine edge direction information of each edge pixel point according to the binary image, and determine a first edge distribution map according to the edge direction information;

[0097] Setting a sliding window on the first edge distribution map, determining a central pixel point of the sliding image included in the sliding window, performing a difference operation between the coordinates of other pixel points in the sliding image except the central pixel point and the coordinates of the central pixel point to obtain a centralized coordinate;

[0098] According to the centralized coordinates and the edge direction information of the central pixel, an association relationship between other pixel points and the central pixel point is established; and a second edge distribution map is determined according to the association relationship.

[0099] The working principle of the above technical solution is as follows: the target image is scored based on several image quality evaluation indicators to obtain corresponding single-item scores; a comprehensive score is calculated based on the single-item scores and preset evaluation weights of several image quality evaluation indicators to obtain a comprehensive score, and it is determined whether it is less than the preset score; when it is determined that the comprehensive score is less than the preset score, the target image is analyzed to determine a local image including texture features and smooth features; the local image is matched with a preset local image in a preset image library, and correction parameters are determined according to the matching results; the local image is denoised and smoothed according to the correction parameters to obtain a corrected image; the corrected image is processed based on an unsharp masking algorithm The method is used to obtain an edge enhanced image; the edge enhanced image is subjected to edge extraction processing based on the Canny operator, and a binary image is generated; the edge trend information of each edge pixel point is determined according to the binary image, and a first edge distribution map is determined according to the edge trend information; a sliding window is set on the first edge distribution map, a central pixel point of the sliding image included in the sliding window is determined, and the coordinates of other pixel points in the sliding image except the central pixel point are subtracted from the coordinates of the central pixel point to obtain a centralized coordinate; according to the centralized coordinate and the edge trend information of the central pixel point, an association relationship between other pixel points and the central pixel point is established; and a second edge distribution map is determined according to the association relationship.

[0100] The beneficial effects of the above technical solution are as follows: when it is determined that the image quality of the target image is low, such as when it contains more image noise, the target image is subjected to denoising and smoothing processing to obtain a corrected image, the image quality of the input image is ensured, the interference of noise is avoided, and the edge features are enhanced based on the unsharp mask algorithm, but the residual noise is also enhanced, so the edge enhancement image is subjected to edge extraction processing based on the Canny operator, and a binary image is generated; the edge direction information of each edge pixel is determined according to the binary image, and the first edge distribution map is determined according to the edge direction information. The effective information is accurately determined, the influence of residual noise is reduced, and the image quality is improved. A sliding window is set on the first edge distribution map to establish the association relationship between other pixel points and the central pixel point; the second edge distribution map is determined according to the association relationship. The correlation of the edge direction information of each pixel point is improved, so that the coordinates of the pixel points in the window can be reduced in dimension, the operation efficiency of the convolution operation of the deep convolutional neural network can be improved, and the recognition accuracy based on the deep convolutional neural network can be improved.

[0101] In one embodiment, determining edge direction information of each edge pixel point according to the binary image includes:

[0102] Select a pixel point with a gray value of 1 in the binary image, and obtain the Hessian matrix of the position of the pixel point; determine the edge direction information of the pixel point according to the Hessian matrix, and then obtain the edge direction information of all pixels with a gray value of 1;

[0103] According to the acquired edge direction information of all pixels with a gray value of 1, the edge direction information of all pixels with a gray value of 0 is determined based on a linear difference algorithm.

[0104] The beneficial effects of the above technical solution are: facilitating accurate determination of edge direction information of each edge pixel point, and improving the accuracy of the determined first edge distribution map.

[0105] According to some embodiments of the present invention, the deep convolutional neural network includes 2 convolutional layers and a computing layer, 2 fully connected layers, 1 input layer and 1 output layer; the sigmoid function is used as the activation function in the convolutional layer; the computing layer is connected to the convolutional layer to calculate the local average and feature extraction; the fully connected layer is used to connect the features determined by the computing layer; the output layer is used to output the final image features.

[0106] According to some embodiments of the present invention, the image features include contour features.

[0107] According to some embodiments of the present invention, matching the image features with preset image features in a database, and determining category information according to the matching results, includes:

[0108] The matching degrees of the image features and the preset image features in the database are calculated respectively, the preset image feature with the greatest matching degree is determined, and the bird information corresponding to the preset image feature with the greatest matching degree is determined as the category information corresponding to the initial image.

[0109] According to some embodiments of the present invention, before inputting the target image into the deep convolutional neural network for convolution operation, the method further includes:

[0110] Input the target image into a U-Net model, perform connected domain analysis on the target image, and obtain a plurality of connected domain images;

[0111] Acquire the areas of several connected domain images, and compare them with preset areas respectively, and use the connected domain images with areas larger than the preset areas as the images to be processed;

[0112] Determine the output probability of all pixels on the image to be processed based on the U-Net model and perform adaptive processing based on preset rules to obtain the corrected probability and compare it with the preset probability;

[0113] The pixel points whose correction probability is greater than the preset probability are taken as reserved pixel points, and a reserved image is obtained according to the reserved pixel points;

[0114] Performing histogram equalization processing on the retained image to obtain an enhanced image;

[0115] Performing threshold processing on the enhanced image based on the maximum inter-class variance method to obtain a grayscale segmentation threshold for image segmentation of the enhanced image;

[0116] Performing binarization processing according to the grayscale segmentation threshold and the grayscale value of each pixel in the enhanced image, and determining the posture of the bird according to the binarization processing result;

[0117] The enhanced image marked with the bird's posture is input into the deep convolutional neural network for convolution operation.

[0118] The working principle and beneficial effects of the above technical solution are as follows: the target image is input into the U-Net model, and a connected domain analysis is performed on the target image to obtain several connected domain images; the areas of several connected domain images are obtained, and compared with the preset areas respectively, and the connected domain images with an area larger than the preset area are used as the images to be processed; the images to be processed are composed of foreground images and part of the noise images. The output probability of all pixels on the image to be processed is determined based on the U-Net model and adaptively processed based on preset rules to obtain the correction probability, and compared with the preset probability; the pixels with a correction probability greater than the preset probability are used as the reserved pixels, and the reserved image is obtained according to the reserved pixels; the reserved pixels in the image to be processed are accurately determined, and then the reserved image, i.e., the foreground image, is obtained. The reserved image is subjected to histogram equalization processing to obtain an enhanced image; it is convenient to expand the grayscale level of the reserved image, that is, the grayscale level of the reserved image is expanded from the original basis to 0 to 256 to enhance the image contrast. The enhanced image is subjected to threshold processing based on the maximum inter-class variance method to obtain a grayscale segmentation threshold for image segmentation of the enhanced image; binarization is performed based on the grayscale segmentation threshold and the grayscale value of each pixel in the enhanced image, and the posture of the bird is determined based on the binarization result; the enhanced image marked with the posture of the bird is input into a deep convolutional neural network for convolution operation. This facilitates accurate determination of the posture of the bird, and further facilitates improving the accuracy of the output results based on the deep convolutional neural network. Adaptive processing is performed based on preset rules to facilitate improving the precision and accuracy of the retained pixels.

[0119] In one embodiment, determining the output probabilities of all pixels on the image to be processed based on the U-Net model and performing adaptive processing based on preset rules to obtain the corrected probabilities includes:

[0120]

[0121] Among them, p ij' is the correction probability of the jth pixel on the i-th image to be processed; p ij is the probability of the jth pixel on the i-th image to be processed based on the output of the U-Net model; s i is the area of ​​the i-th image to be processed; m is the number of pixels included in the i-th image to be processed; n is the number of images to be processed.

[0122] The working principle and beneficial effects of the above technical solution are as follows: the correction probability after adaptive processing is accurately calculated, thereby improving the accuracy of judging the correction probability and the preset probability. The precision and accuracy of determining the retained pixels are improved.

[0123] like Figure 3 As shown, the second embodiment of the present invention proposes an intelligent model based on an AI chip to support intelligent identification of birds, including:

[0124] A collection module, used for collecting initial images including birds;

[0125] A processing module, used for normalizing the size of the initial image to obtain a target image and inputting the target image into a deep convolutional neural network for convolution operation to obtain image features;

[0126] The matching module is used to match the image features with preset image features in the database and determine the category information according to the matching results.

[0127] The working principle of the above technical solution is: an acquisition module is used to acquire an initial image including birds; a processing module is used to standardize the size of the initial image, obtain a target image and input it into a deep convolutional neural network for convolution operation to obtain image features; a matching module is used to match the image features with preset image features in a database, and determine category information based on the matching results.

[0128] The beneficial effects of the above technical solution are as follows: the size of the initial image is standardized to meet the input parameter requirements of the deep convolutional neural network for convolution operations, which facilitates obtaining accurate image features, and then determines the category information based on the matching results, thereby facilitating the intelligent recognition of birds and improving recognition efficiency and accuracy.

[0129] According to some embodiments of the present invention, the further comprising:

[0130] Regulation modules for:

[0131] Before the processing module inputs the target image into the deep convolutional neural network for convolution operation, grayscale processing is performed on the target image to obtain a grayscale image;

[0132] Obtaining the grayscale values ​​of the pixels in the grayscale image, screening out the maximum grayscale value and the minimum grayscale value, and calculating the grayscale mean and standard deviation;

[0133] Calculating the addition value and subtraction value of the grayscale mean and standard deviation;

[0134] Compare the maximum grayscale value with the added value, and determine the larger value as the first calculated value;

[0135] Compare the minimum grayscale value with the subtraction value, and determine the smaller value as the second calculated value;

[0136] Calculating a ratio of the second calculated value to the first calculated value, and determining whether the ratio is less than a preset ratio;

[0137] When it is determined that the ratio is less than a preset ratio, Gaussian blur processing is performed on the grayscale image to determine the grayscale value of each pixel in the grayscale image after the Gaussian blur processing;

[0138] Determine the grayscale difference of each pixel in the grayscale image before and after Gaussian blur processing, determine the adjustment parameter of the corresponding pixel according to the grayscale difference, and perform adjustment processing on the corresponding pixel according to the adjustment parameter.

[0139] The working principle of the above technical solution: the adjustment module is used to: before the processing module inputs the target image into the deep convolutional neural network for convolution operation, grayscale processing is performed on the target image to obtain a grayscale image; the grayscale values ​​of the pixels in the grayscale image are obtained, the maximum grayscale value and the minimum grayscale value are screened out, and the grayscale mean and standard deviation are calculated; the addition value and subtraction value of the grayscale mean and the standard deviation are calculated; the maximum grayscale value is compared with the addition value, and the larger value is determined as the first calculation value; the minimum grayscale value is compared with the subtraction value, and the smaller value is determined as the second calculation value; the ratio of the second calculation value to the first calculation value is calculated, and it is determined whether it is less than a preset ratio; when it is determined that the ratio is less than the preset ratio, Gaussian blur processing is performed on the grayscale image to determine the grayscale value of each pixel in the grayscale image after Gaussian blur processing; the grayscale difference of each pixel in the grayscale image before and after Gaussian blur processing is determined, and the adjustment parameters of the corresponding pixel are determined according to the grayscale difference, and the corresponding pixel is adjusted according to the adjustment parameters.

[0140] The beneficial effect of the above technical solution is: when it is determined that the clarity of the grayscale image is low, clarity adjustment processing is performed to ensure the clarity of the input image, thereby facilitating improving the recognition accuracy based on the deep convolutional neural network.

[0141] According to some embodiments of the present invention, the further comprising:

[0142] The correction module is used to score the target image based on several image quality evaluation indicators before the processing module inputs the target image into the deep convolutional neural network for convolution operation, and obtain the corresponding single score; calculate according to the single score and the preset evaluation weights of several image quality evaluation indicators to obtain a comprehensive score, and determine whether it is less than the preset score;

[0143] When it is determined that the comprehensive score is less than the preset score, the target image is analyzed to determine a local image including texture features and smooth features;

[0144] Matching the local image with a preset local image in a preset image library, and determining a correction parameter according to the matching result;

[0145] Performing noise reduction and smoothing processing on the local image according to the correction parameters to obtain a corrected image;

[0146] Processing the corrected image based on an unsharp mask algorithm to obtain an edge enhanced image;

[0147] Performing edge extraction processing on the edge enhanced image based on the Canny operator and generating a binary image;

[0148] Determine edge direction information of each edge pixel point according to the binary image, and determine a first edge distribution map according to the edge direction information;

[0149] Setting a sliding window on the first edge distribution map, determining a central pixel point of the sliding image included in the sliding window, performing a difference operation between the coordinates of other pixel points in the sliding image except the central pixel point and the coordinates of the central pixel point to obtain a centralized coordinate;

[0150] According to the centralized coordinates and the edge direction information of the central pixel, an association relationship between other pixel points and the central pixel point is established; and a second edge distribution map is determined according to the association relationship.

[0151] The beneficial effects of the above technical solution are as follows: when it is determined that the image quality of the target image is low, such as when it contains more image noise, the target image is subjected to denoising and smoothing processing to obtain a corrected image, the image quality of the input image is ensured, the interference of noise is avoided, and the edge features are enhanced based on the unsharp mask algorithm, but the residual noise is also enhanced, so the edge enhancement image is subjected to edge extraction processing based on the Canny operator, and a binary image is generated; the edge direction information of each edge pixel is determined according to the binary image, and the first edge distribution map is determined according to the edge direction information. The effective information is accurately determined, the influence of residual noise is reduced, and the image quality is improved. A sliding window is set on the first edge distribution map to establish the association relationship between other pixel points and the central pixel point; the second edge distribution map is determined according to the association relationship. The correlation of the edge direction information of each pixel point is improved, so that the coordinates of the pixel points in the window can be reduced in dimension, the operation efficiency of the convolution operation of the deep convolutional neural network can be improved, and the recognition accuracy based on the deep convolutional neural network can be improved.

[0152] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. An intelligent algorithm based on AI chip to support intelligent bird identification. It is characterized in that include: The acquisition includes initial images of the birds; The size of the initial image is normalized to obtain a target image and the target image is input into a deep convolutional neural network for convolution operation to obtain image features; Matching the image features with preset image features in a database, and determining category information according to the matching results; Before the target image is input into the deep convolutional neural network for convolution operation, it also includes: Performing grayscale processing on the target image to obtain a grayscale image; Obtaining the grayscale values ​​of the pixels in the grayscale image, screening out the maximum grayscale value and the minimum grayscale value, and calculating the grayscale mean and standard deviation; Calculating the addition value and subtraction value of the grayscale mean and standard deviation; Compare the maximum grayscale value with the added value, and determine the larger value as the first calculated value; Compare the minimum grayscale value with the subtraction value, and determine the smaller value as the second calculated value; Calculating a ratio of the second calculated value to the first calculated value, and determining whether the ratio is less than a preset ratio; When it is determined that the ratio is less than a preset ratio, Gaussian blur processing is performed on the grayscale image to determine the grayscale value of each pixel in the grayscale image after the Gaussian blur processing; Determine the grayscale difference of each pixel in the grayscale image before and after Gaussian blur processing, determine the adjustment parameter of the corresponding pixel according to the grayscale difference, and perform adjustment processing on the corresponding pixel according to the adjustment parameter; Before the target image is input into the deep convolutional neural network for convolution operation, it also includes: Input the target image into a U-Net model, perform connected domain analysis on the target image, and obtain a plurality of connected domain images; Acquire the areas of several connected domain images, and compare them with preset areas respectively, and use the connected domain images with areas larger than the preset areas as the images to be processed; Determine the output probability of all pixels on the image to be processed based on the U-Net model and perform adaptive processing based on preset rules to obtain the corrected probability and compare it with the preset probability; The pixel points whose correction probability is greater than the preset probability are taken as reserved pixel points, and a reserved image is obtained according to the reserved pixel points; Performing histogram equalization processing on the retained image to obtain an enhanced image; Performing threshold processing on the enhanced image based on the maximum inter-class variance method to obtain a grayscale segmentation threshold for image segmentation of the enhanced image; Performing binarization processing according to the grayscale segmentation threshold and the grayscale value of each pixel in the enhanced image, and determining the posture of the bird according to the binarization processing result; The enhanced image marked with the bird's posture is input into the deep convolutional neural network for convolution operation.

2. The intelligent algorithm for supporting intelligent bird identification based on an AI chip as claimed in claim 1, It is characterized in that Before the target image is input into the deep convolutional neural network for convolution operation, it also includes: Scoring the target image based on a number of image quality evaluation indicators to obtain a corresponding single score; calculating according to the single score and preset evaluation weights of the number of image quality evaluation indicators to obtain a comprehensive score, and determining whether the comprehensive score is less than the preset score; When it is determined that the comprehensive score is less than the preset score, the target image is analyzed to determine a local image including texture features and smooth features; Matching the local image with a preset local image in a preset image library, and determining a correction parameter according to the matching result; Performing noise reduction and smoothing processing on the local image according to the correction parameters to obtain a corrected image; Processing the corrected image based on an unsharp mask algorithm to obtain an edge enhanced image; Performing edge extraction processing on the edge enhanced image based on the Canny operator and generating a binary image; Determine edge direction information of each edge pixel point according to the binary image, and determine a first edge distribution map according to the edge direction information; Setting a sliding window on the first edge distribution map, determining a central pixel point of the sliding image included in the sliding window, performing a difference operation between the coordinates of other pixel points in the sliding image except the central pixel point and the coordinates of the central pixel point to obtain a centralized coordinate; According to the centralized coordinates and the edge direction information of the central pixel, an association relationship between other pixel points and the central pixel point is established; and a second edge distribution map is determined according to the association relationship.

3. The intelligent algorithm for supporting intelligent bird identification based on an AI chip as claimed in claim 1, It is characterized in that The deep convolutional neural network includes 2 convolutional layers and a computing layer, 2 fully connected layers, 1 input layer and 1 output layer; the sigmoid function is used as the activation function in the convolutional layer; the computing layer is connected to the convolutional layer to calculate the local average and feature extraction; the fully connected layer is used to connect the features determined by the computing layer; the output layer is used to output the final image features.

4. The intelligent algorithm for supporting intelligent bird identification based on an AI chip as claimed in claim 1, It is characterized in that The image features include contour features.

5. The intelligent algorithm for supporting intelligent bird identification based on an AI chip as claimed in claim 1, It is characterized in that Matching the image features with preset image features in a database, and determining category information according to the matching results, including: The matching degrees of the image features and the preset image features in the database are calculated respectively, the preset image feature with the greatest matching degree is determined, and the bird information corresponding to the preset image feature with the greatest matching degree is determined as the category information corresponding to the initial image.

6. An intelligent model based on AI chip to support intelligent bird identification. It is characterized in that include: A collection module, used for collecting initial images including birds; A processing module, used for normalizing the size of the initial image to obtain a target image and inputting the target image into a deep convolutional neural network for convolution operation to obtain image features; A matching module, used to match the image features with preset image features in a database, and determine category information according to the matching results; Also includes: Regulation modules for: Before the processing module inputs the target image into the deep convolutional neural network for convolution operation, grayscale processing is performed on the target image to obtain a grayscale image; Obtaining the grayscale values ​​of the pixels in the grayscale image, screening out the maximum grayscale value and the minimum grayscale value, and calculating the grayscale mean and standard deviation; Calculating the addition value and subtraction value of the grayscale mean and standard deviation; Compare the maximum grayscale value with the added value, and determine the larger value as the first calculated value; Compare the minimum grayscale value with the subtraction value, and determine the smaller value as the second calculated value; Calculating a ratio of the second calculated value to the first calculated value, and determining whether the ratio is less than a preset ratio; When it is determined that the ratio is less than a preset ratio, Gaussian blur processing is performed on the grayscale image to determine the grayscale value of each pixel in the grayscale image after the Gaussian blur processing; Determine the grayscale difference of each pixel in the grayscale image before and after Gaussian blur processing, determine the adjustment parameter of the corresponding pixel according to the grayscale difference, and perform adjustment processing on the corresponding pixel according to the adjustment parameter; Before the target image is input into the deep convolutional neural network for convolution operation, it also includes: Input the target image into a U-Net model, perform connected domain analysis on the target image, and obtain a plurality of connected domain images; Acquire the areas of several connected domain images, and compare them with preset areas respectively, and use the connected domain images with areas larger than the preset areas as the images to be processed; Determine the output probability of all pixels on the image to be processed based on the U-Net model and perform adaptive processing based on preset rules to obtain the corrected probability and compare it with the preset probability; The pixel points whose correction probability is greater than the preset probability are taken as reserved pixel points, and a reserved image is obtained according to the reserved pixel points; Performing histogram equalization processing on the retained image to obtain an enhanced image; Performing threshold processing on the enhanced image based on the maximum inter-class variance method to obtain a grayscale segmentation threshold for image segmentation of the enhanced image; Performing binarization processing according to the grayscale segmentation threshold and the grayscale value of each pixel in the enhanced image, and determining the posture of the bird according to the binarization processing result; The enhanced image marked with the bird's posture is input into the deep convolutional neural network for convolution operation.

7. The intelligent model for supporting intelligent bird identification based on an AI chip as claimed in claim 6, It is characterized in that Also includes: The correction module is used to score the target image based on several image quality evaluation indicators before the processing module inputs the target image into the deep convolutional neural network for convolution operation, and obtain the corresponding single score; calculate according to the single score and the preset evaluation weights of several image quality evaluation indicators to obtain a comprehensive score, and determine whether it is less than the preset score; When it is determined that the comprehensive score is less than the preset score, the target image is analyzed to determine a local image including texture features and smooth features; Matching the local image with a preset local image in a preset image library, and determining a correction parameter according to the matching result; Performing noise reduction and smoothing processing on the local image according to the correction parameters to obtain a corrected image; Processing the corrected image based on an unsharp mask algorithm to obtain an edge enhanced image; Performing edge extraction processing on the edge enhanced image based on the Canny operator and generating a binary image; Determine edge direction information of each edge pixel point according to the binary image, and determine a first edge distribution map according to the edge direction information; Setting a sliding window on the first edge distribution map, determining a central pixel point of the sliding image included in the sliding window, performing a difference operation between the coordinates of other pixel points in the sliding image except the central pixel point and the coordinates of the central pixel point to obtain a centralized coordinate; According to the centralized coordinates and the edge direction information of the central pixel, an association relationship between other pixel points and the central pixel point is established; and a second edge distribution map is determined according to the association relationship.

Images