Fry quantity estimation method based on multi-scale semantic fusion density estimation network

Abstract

The invention relates to the technical field of image processing, in particular to a fry quantity estimation method based on a multi-scale semantic fusion density estimation network, which comprises the following steps of: firstly, carrying out grid division on a fish pond, and obtaining a first fry image to be detected, illumination intensity and area of each grid region; constructing a fry quantity estimation model to process the to-be-detected image, obtaining a gray-level co-occurrence matrix of the to-be-detected fry image through image recognition, obtaining texture features, further obtaining a first turbidity coefficient, and determining different denoising modes according to the first turbidity coefficient; the feature extraction layer carries out processing through a large-scale feature extraction unit, a small-scale feature extraction unit and an attention mechanism; the feature fusion layer obtains a first fry fusion feature map; the feature analysis layer generates a first fry density map by using the first fry fusion feature map to obtain a first fry number, and corrects the first fry number according to illumination intensity to obtain a second fry number; the method can effectively improve the accuracy of fry quantity estimation.

Description

Technical Field

[0001] The invention relates to the technical field of image processing, and in particular to a method for estimating the number of fry based on a multi-scale semantic fusion density estimation network. Background Art

[0002] In the process of aquaculture, the estimation of the number of fry is an important part of production management. At present, most traditional methods of estimating the number of fry estimate fish through single-scale feature analysis, and do not fully consider the impact of environmental influences on the estimation of the number of fry, which easily leads to limited accuracy of the estimation of the number of fry.

[0003] Therefore, a fry quantity estimation method based on multi-scale semantic fusion density estimation network was proposed. Summary of the invention

[0004] The purpose of the present invention is to provide a method for estimating the number of fry based on a multi-scale semantic fusion density estimation network. The present invention relates to the field of image processing technology, and specifically to a method for estimating the number of fry based on a multi-scale semantic fusion density estimation network. The method comprises: firstly, the present invention divides a fish pond into grids, obtains a first fry image to be detected, light intensity and area of ​​each grid area; constructs a fry quantity estimation model to process the image to be detected, obtains a grayscale co-occurrence matrix of the fry image to be detected by image recognition, obtains texture features, and then obtains a first turbidity coefficient, and determines different denoising methods according to the first turbidity coefficient; a feature extraction layer is processed by a large-scale feature extraction unit, a small-scale feature extraction unit and an attention mechanism; a feature fusion layer obtains a first fry fusion feature map; a feature analysis layer generates a first fry density map using the first fry fusion feature map, obtains a first fry number, and obtains a second fry number according to light intensity correction; the method can effectively improve the accuracy of fry quantity estimation.

[0005] To achieve the above object, the present invention provides the following technical solutions: The method for estimating the number of fry based on a multi-scale semantic fusion density estimation network includes: S1. Grid the fish pond to obtain grid areas of each fish pond, obtain the first image of the fry to be detected, the second light intensity and the third area of ​​each fish pond grid area; S2. Construct a fish fry quantity estimation model to identify the first fish fry image to be detected, the fish fry quantity estimation model includes an input layer, a preprocessing layer, a feature extraction layer, a feature fusion layer, a feature analysis layer and an output layer; the preprocessing layer performs image recognition on the first fish fry image to be detected to obtain a first turbidity coefficient; determine different denoising methods according to the first turbidity coefficient and a preset turbidity coefficient threshold set; and perform multi-scale image segmentation on the first fish fry image to be detected to obtain multiple first large-scale fish fry detection images and second small-scale fish fry detection images; S3. The feature extraction layer extracts features of the first large-scale fry detection image and the second small-scale fry detection image through a large-scale feature extraction unit and a small-scale feature extraction unit, respectively, and processes them through the SE attention mechanism to obtain a first large-scale feature map and a second small-scale feature map respectively; the feature fusion layer is used to fuse the first large-scale feature map and the second small-scale feature map to obtain a first fry fusion feature map; S4. The feature analysis layer obtains a first fry density map according to the first fry fusion feature map; obtains a first fry number according to the first fry density map and the third area; and obtains a second fry number according to the first fry number, the second light intensity and the first turbidity coefficient.

[0006] Preferably, the preprocessing layer performs image recognition on the first fry image to be detected to obtain the first turbidity coefficient specifically includes: S10. graying the first fry image to be detected, and calculating the gray level co-occurrence matrix of the first fry image to be detected; S20. Extracting texture features according to the gray level co-occurrence matrix, wherein the texture features include contrast, uniformity and energy; S30. Obtain a first turbidity coefficient according to the texture feature.

[0007] Preferably, the determining of different denoising methods according to the first turbidity coefficient and the preset turbidity coefficient threshold set specifically includes: S100. Determine the turbidity level of the fish pond grid area according to the first turbidity coefficient and a preset turbidity coefficient threshold set, wherein the turbidity level includes light turbidity, moderate turbidity and heavy turbidity; S200. De-noise the slightly turbid fish pond grid area by Gaussian filtering; de-noise the moderately turbid fish pond grid area by bilateral filtering; de-noise the heavily turbid fish pond grid area by deep learning denoising.

[0008] Preferably, the preset turbidity coefficient threshold set includes a first turbidity coefficient threshold and a second turbidity coefficient threshold; the specific process of obtaining the turbidity level is: ; in, Indicates turbidity level; Indicates mild turbidity; represents the first turbidity coefficient; represents the first turbidity coefficient threshold; represents the second turbidity coefficient threshold; Indicates moderate turbidity; Indicates severe turbidity.

[0009] Preferably, the fry quantity estimation model comprises an input layer, a preprocessing layer, a feature extraction layer, a feature fusion layer, a feature analysis layer and an output layer; The preprocessing layer performs image recognition on the first fry image to be detected to obtain a first turbidity coefficient, and determines different denoising methods according to the first turbidity coefficient; and performs multi-scale image segmentation on the first fry image to be detected through U-Net to obtain a plurality of first large-scale fry detection images and second small-scale fry detection images; The feature extraction layer extracts features of the first large-scale fish fry detection image and the second small-scale fish fry detection image through a large-scale feature extraction unit and a small-scale feature extraction unit, respectively, and processes them through an SE attention mechanism to obtain a first large-scale feature map and a second small-scale feature map, respectively; The feature fusion layer is used to fuse the first large-scale feature map and the second small-scale feature map to obtain a first fry fusion feature map; The feature analysis layer performs channel dimension reduction on the first fry fusion feature map through the first convolution kernel, and sequentially passes through the second convolution kernel and the third convolution kernel to obtain a first fry density map; obtains a first fry number according to the first fry density map and the third area; and obtains a second fry number according to the first fry number, the second light intensity and the first turbidity coefficient; The output layer is used to output the second number of fry.

[0010] Preferably, the first number of fry is: ; in, Indicates the number of first fry; represents the third area; Indicates the number of fish pond grid areas; represents the total number of pixels in the first fry density map; Represents the sum of pixel values ​​of the first fry density map.

[0011] Preferably, the second number of fry is: ; in, Indicates the second number of fry; Indicates the number of first fry; Indicates the influence weight of light intensity; Indicates the average light intensity in the fish pond; Indicates the light intensity reference value.

[0012] Preferably, the average light intensity of the fish pond is obtained based on the second light intensity and the number of grid areas in the fish pond.

[0013] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention uses image recognition to identify the grayscale co-occurrence matrix of the image to be detected of the fry, further obtains the texture features of the image to be detected of the fry, obtains the first turbidity coefficient according to the texture features, and determines the turbidity level of the fish pond grid area according to the first turbidity coefficient and the preset turbidity coefficient threshold, and determines different denoising methods according to the turbidity level; this method can effectively avoid the influence of fish pond turbidity on the estimation of the number of fry, provides a good data image basis for the later model analysis, and further effectively improves the accuracy of the estimation of the number of fry.

[0014] 2. The present invention identifies the fry images to be detected through a fry quantity estimation model, divides the fry images to be detected in each fish pond grid area into large-scale images and small-scale images through image segmentation, and extracts features through a large-scale feature extraction unit and a small-scale feature extraction unit in a feature extraction layer, and processes the feature through an SE attention mechanism to obtain a first large-scale feature map and a second small-scale feature map; then, the first large-scale feature map and the second small-scale feature map are fused according to a feature fusion layer to obtain a first fry fusion feature map; a first fry density map is obtained according to the first fry fusion feature map, and then a first fry number is obtained. The method can effectively integrate the multi-scale features of the fry images to be detected through the fry quantity estimation model, effectively improve the recognition ability of the fry images to be detected, and thus effectively improve the accuracy of fry quantity estimation.

[0015] 3. The present invention analyzes the light intensity of each fish pond grid area and the first number of fry to further obtain the second number of fry. Through effective correction, the influence of light intensity on the estimation of the number of fry can be avoided, thereby further improving the accuracy of the estimation of the number of fry. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A schematic flow chart of a method for estimating the number of fry based on a multi-scale semantic fusion density estimation network provided by an embodiment of the present invention; Figure 2 A schematic diagram of a denoising process provided by an embodiment of the present invention; Figure 3 A schematic diagram of the structure of a fry quantity estimation model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.

[0018] Embodiment 1 In order to improve the accuracy of estimating the number of fry in fish pond A, a method of estimating the number of fry based on a multi-scale semantic fusion density estimation network was applied. Reference Figure 1 The schematic flow chart of the method for estimating the number of fry based on the multi-scale semantic fusion density estimation network provided by the embodiment of the present invention specifically includes: S1. Grid the fish pond to obtain grid areas of each fish pond, obtain the first image of the fry to be detected, the second light intensity and the third area of ​​each fish pond grid area; Furthermore, in this embodiment, the number of fish pond grid regions is N; the third area represents the area of ​​each fish pond grid region; S2. Construct a fish fry quantity estimation model to identify the first fish fry image to be detected, the fish fry quantity estimation model includes an input layer, a preprocessing layer, a feature extraction layer, a feature fusion layer, a feature analysis layer and an output layer; the preprocessing layer performs image recognition on the first fish fry image to be detected to obtain a first turbidity coefficient; determine different denoising methods according to the first turbidity coefficient and a preset turbidity coefficient threshold set; and perform multi-scale image segmentation on the first fish fry image to be detected to obtain multiple first large-scale fish fry detection images and second small-scale fish fry detection images; Furthermore, the fry quantity estimation model includes an input layer, a preprocessing layer, a feature extraction layer, a feature fusion layer, a feature analysis layer and an output layer; Figure 3 A schematic diagram of the structure of a fish fry quantity estimation model provided by an embodiment of the present invention; The preprocessing layer performs image recognition on the first fry image to be detected to obtain a first turbidity coefficient, and determines different denoising methods according to the first turbidity coefficient; and performs multi-scale image segmentation on the first fry image to be detected through U-Net to obtain a plurality of first large-scale fry detection images and second small-scale fry detection images; The feature extraction layer extracts features of the first large-scale fish fry detection image and the second small-scale fish fry detection image through a large-scale feature extraction unit and a small-scale feature extraction unit, respectively, and processes them through an SE attention mechanism to obtain a first large-scale feature map and a second small-scale feature map, respectively; The feature fusion layer is used to fuse the first large-scale feature map and the second small-scale feature map to obtain a first fry fusion feature map; The feature analysis layer performs channel dimension reduction on the first fry fusion feature map through the first convolution kernel, and sequentially passes through the second convolution kernel and the third convolution kernel to obtain a first fry density map; the size of the first convolution kernel is 1×1; the size of the second convolution kernel is 3×3; the size of the third convolution kernel is 1×1; according to the first fry density map and the third area, the first fry number is obtained; and according to the first fry number, the second light intensity and the first turbidity coefficient, the second fry number is obtained; The output layer is used to output the second number of fry.

[0019] Furthermore, the preprocessing layer performs image recognition on the first fry image to be detected to obtain the first turbidity coefficient specifically including: S10. graying the first fry image to be detected, and calculating the gray level co-occurrence matrix of the first fry image to be detected; S20. Extracting texture features according to the gray level co-occurrence matrix, wherein the texture features include contrast, uniformity and energy; S30. Obtaining a first turbidity coefficient according to the texture feature; Furthermore, the first turbidity coefficient is: ; in, represents the first turbidity coefficient; Indicates the contrast influence weight; Indicates contrast, reflecting grayscale differences; represents the uniformity impact weight; It indicates uniformity and reflects the regularity of texture; represents the energy impact weight; Indicates energy, which reflects the smoothness of the texture; Furthermore, the determining of different denoising methods according to the first turbidity coefficient and the preset turbidity coefficient threshold set specifically includes: S100. Determine the turbidity level of the fish pond grid area according to the first turbidity coefficient and a preset turbidity coefficient threshold set, wherein the turbidity level includes light turbidity, moderate turbidity and heavy turbidity; S200. De-noise the slightly turbid fish pond grid area by Gaussian filtering; de-noise the moderately turbid fish pond grid area by bilateral filtering; de-noise the heavily turbid fish pond grid area by deep learning denoising.

[0020] Furthermore, the preset turbidity coefficient threshold set includes a first turbidity coefficient threshold and a second turbidity coefficient threshold; the specific acquisition process of the turbidity level is: ; in, Indicates turbidity level; Indicates mild turbidity; represents the first turbidity coefficient; represents the first turbidity coefficient threshold; represents the second turbidity coefficient threshold; Indicates moderate turbidity; Indicates severe turbidity; The first turbidity coefficient threshold and the second turbidity coefficient threshold are obtained by training with historical fish pond image data; The specific denoising process diagram is as follows Figure 2 As shown; This embodiment uses image recognition to obtain the grayscale co-occurrence matrix of the image of the fry to be detected, further obtains the texture features of the image of the fry to be detected, obtains the first turbidity coefficient according to the texture features, and determines the turbidity level of the fish pond grid area according to the first turbidity coefficient and a preset turbidity coefficient threshold, and determines different denoising methods according to the turbidity level; this method can effectively avoid the influence of the turbidity of the fish pond on the estimation of the number of fry, provides a good data image foundation for the subsequent model analysis, and further effectively improves the accuracy of the estimation of the number of fry.

[0021] S3. The feature extraction layer extracts features of the first large-scale fry detection image and the second small-scale fry detection image through a large-scale feature extraction unit and a small-scale feature extraction unit, respectively, and processes them through the SE attention mechanism to obtain a first large-scale feature map and a second small-scale feature map respectively; the feature fusion layer is used to fuse the first large-scale feature map and the second small-scale feature map to obtain a first fry fusion feature map; S4. The feature analysis layer obtains a first fry density map according to the first fry fusion feature map; obtains a first fry number according to the first fry density map and the third area; and obtains a second fry number according to the first fry number, the second light intensity and the first turbidity coefficient; Further, the first number of fry is: ; in, Indicates the number of first fry; represents the third area; Indicates the number of fish pond grid areas; represents the total number of pixels in the first fry density map; Represents the sum of pixel values ​​of the first fry density map.

[0022] In this embodiment, the fry quantity estimation model is used to identify the fry images to be detected, and the fry images to be detected in each fish pond grid area are divided into large-scale images and small-scale images through image segmentation. The feature extraction layer is distributed through a large-scale feature extraction unit and a small-scale feature extraction unit to extract features, and is processed through the SE attention mechanism to obtain a first large-scale feature map and a second small-scale feature map; then, the first large-scale feature map and the second small-scale feature map are fused according to the feature fusion layer to obtain a first fry fusion feature map; a first fry density map is obtained according to the first fry fusion feature map, and then the first fry number is obtained. This method can effectively integrate the multi-scale features of the fry images to be detected through the fry quantity estimation model, effectively improve the recognition ability of the fry images to be detected, and thus can effectively improve the accuracy of the fry quantity estimation.

[0023] Further, the second number of fry is: ; in, Indicates the second number of fry; Indicates the number of first fry; Indicates the influence weight of light intensity; Indicates the average light intensity in the fish pond; Indicates the light intensity reference value.

[0024] Further, the average light intensity of the fish pond is obtained according to the second light intensity and the number of grid areas of the fish pond; The average light intensity of the fish pond is: ; in, Indicates the average light intensity in the fish pond; Indicates the number of fish pond grid areas; represents the second light intensity; This embodiment further obtains the second number of fry by analyzing the light intensity of each fish pond grid area and the first number of fry. Through effective correction, the influence of light intensity on the estimation of the number of fry can be avoided, thereby further improving the accuracy of the estimation of the number of fry.

[0025] In this embodiment, the fish pond is firstly divided into grids, and the first image of fry to be detected, the light intensity and the area of ​​each grid area are obtained; a fry quantity estimation model is constructed to process the image to be detected, and the grayscale co-occurrence matrix of the image of fry to be detected is obtained through image recognition, and the texture feature is obtained, and then the first turbidity coefficient is obtained, and different denoising methods are determined according to the first turbidity coefficient; the feature extraction layer is processed by a large-scale feature extraction unit, a small-scale feature extraction unit and an attention mechanism; the feature fusion layer obtains the first fry fusion feature map; the feature analysis layer generates a first fry density map using the first fry fusion feature map, obtains the first fry number, and corrects it according to the light intensity to obtain the second fry number; this method can effectively improve the accuracy of fry quantity estimation.

[0026] In order to verify the effectiveness of the method for estimating the number of fry based on the multi-scale semantic fusion density estimation network provided by the present invention, the accuracy of estimating the number of fry of different methods is compared and distributed, including method one, method two and method three; method one is the method for estimating the number of fry based on the multi-scale semantic fusion density estimation network provided in this embodiment; method two is based on method one without considering denoising; method three is based on method one without considering the influence of light intensity, and the mean accuracy of the estimation of the number of fry in fish pond A by different methods is given, and the specific results are shown in Table 1; Table 1 Mean accuracy of different methods for estimating the number of fry in pond A method Estimated mean accuracy Method 1 96% Method 2 89% Method 3 86% It can be seen from Table 1 that the effectiveness of the method for estimating the number of fry based on the multi-scale semantic fusion density estimation network provided in this embodiment can effectively improve the accuracy of estimating the number of fry.

[0027] Embodiment 2 In order to improve the accuracy of estimating the number of fry in fish pond B, a method of estimating the number of fry based on a multi-scale semantic fusion density estimation network was applied; Reference Figure 1 The schematic flow chart of the method for estimating the number of fry based on the multi-scale semantic fusion density estimation network provided by the embodiment of the present invention specifically includes: S1. Grid the fish pond to obtain each fish pond grid area, obtain the first fry image to be detected, the second light intensity and the third area of ​​each fish pond grid area; further, in this embodiment, the number of fish pond grid areas is N; the third area represents the area of ​​each fish pond grid area; S2. Construct a fish fry quantity estimation model to identify the first fish fry image to be detected, the fish fry quantity estimation model includes an input layer, a preprocessing layer, a feature extraction layer, a feature fusion layer, a feature analysis layer and an output layer; the preprocessing layer performs image recognition on the first fish fry image to be detected to obtain a first turbidity coefficient; determine different denoising methods according to the first turbidity coefficient and a preset turbidity coefficient threshold set; and perform multi-scale image segmentation on the first fish fry image to be detected to obtain multiple first large-scale fish fry detection images and second small-scale fish fry detection images; Furthermore, the fry quantity estimation model includes an input layer, a preprocessing layer, a feature extraction layer, a feature fusion layer, a feature analysis layer and an output layer; Figure 3 A schematic diagram of the structure of a fish fry quantity estimation model provided by an embodiment of the present invention; The preprocessing layer performs image recognition on the first fry image to be detected to obtain a first turbidity coefficient, and determines different denoising methods according to the first turbidity coefficient; and performs multi-scale image segmentation on the first fry image to be detected through U-Net to obtain a plurality of first large-scale fry detection images and second small-scale fry detection images; The feature extraction layer extracts features of the first large-scale fish fry detection image and the second small-scale fish fry detection image through a large-scale feature extraction unit and a small-scale feature extraction unit, respectively, and processes them through an SE attention mechanism to obtain a first large-scale feature map and a second small-scale feature map, respectively; The feature fusion layer is used to fuse the first large-scale feature map and the second small-scale feature map to obtain a first fry fusion feature map; The feature analysis layer performs channel dimension reduction on the first fry fusion feature map through the first convolution kernel, and sequentially passes through the second convolution kernel and the third convolution kernel to obtain a first fry density map; the size of the first convolution kernel is 1×1; the size of the second convolution kernel is 3×3; the size of the third convolution kernel is 1×1; according to the first fry density map and the third area, the first fry number is obtained; and according to the first fry number, the second light intensity and the first turbidity coefficient, the second fry number is obtained; The output layer is used to output the second number of fry.

[0028] Furthermore, the preprocessing layer performs image recognition on the first fry image to be detected to obtain the first turbidity coefficient specifically including: S10. graying the first fry image to be detected, and calculating the gray level co-occurrence matrix of the first fry image to be detected; S20. Extracting texture features according to the gray level co-occurrence matrix, wherein the texture features include contrast, uniformity and energy; S30. Obtaining a first turbidity coefficient according to the texture feature; Furthermore, the first turbidity coefficient is: ; in, represents the first turbidity coefficient; Indicates the contrast influence weight; Indicates contrast, reflecting grayscale differences; represents the uniformity impact weight; It indicates uniformity and reflects the regularity of texture; represents the energy impact weight; Indicates energy, which reflects the smoothness of the texture; Furthermore, the determining of different denoising methods according to the first turbidity coefficient and the preset turbidity coefficient threshold set specifically includes: S100. Determine the turbidity level of the fish pond grid area according to the first turbidity coefficient and a preset turbidity coefficient threshold set, wherein the turbidity level includes light turbidity, moderate turbidity and heavy turbidity; S200. De-noise the slightly turbid fish pond grid area by Gaussian filtering; de-noise the moderately turbid fish pond grid area by bilateral filtering; de-noise the heavily turbid fish pond grid area by deep learning denoising.

[0029] Furthermore, the preset turbidity coefficient threshold set includes a first turbidity coefficient threshold and a second turbidity coefficient threshold; the specific acquisition process of the turbidity level is: ; in, Indicates turbidity level; Indicates mild turbidity; represents the first turbidity coefficient; represents the first turbidity coefficient threshold; represents the second turbidity coefficient threshold; Indicates moderate turbidity; Indicates severe turbidity; The first turbidity coefficient threshold and the second turbidity coefficient threshold are obtained by training with historical fish pond image data; The specific denoising process diagram is as follows Figure 2 As shown; This embodiment uses image recognition to obtain the grayscale co-occurrence matrix of the image of the fry to be detected, further obtains the texture features of the image of the fry to be detected, obtains the first turbidity coefficient according to the texture features, and determines the turbidity level of the fish pond grid area according to the first turbidity coefficient and a preset turbidity coefficient threshold, and determines different denoising methods according to the turbidity level; this method can effectively avoid the influence of the turbidity of the fish pond on the estimation of the number of fry, provides a good data image foundation for the subsequent model analysis, and further effectively improves the accuracy of the estimation of the number of fry.

[0030] S3. The feature extraction layer extracts features of the first large-scale fry detection image and the second small-scale fry detection image through a large-scale feature extraction unit and a small-scale feature extraction unit, respectively, and processes them through the SE attention mechanism to obtain a first large-scale feature map and a second small-scale feature map respectively; the feature fusion layer is used to fuse the first large-scale feature map and the second small-scale feature map to obtain a first fry fusion feature map; S4. The feature analysis layer obtains a first fry density map according to the first fry fusion feature map; obtains a first fry number according to the first fry density map and the third area; and obtains a second fry number according to the first fry number, the second light intensity and the first turbidity coefficient; Further, the first number of fry is: ; in, Indicates the number of first fry; represents the third area; Indicates the number of fish pond grid areas; represents the total number of pixels in the first fry density map; Represents the sum of pixel values ​​of the first fry density map.

[0031] In this embodiment, the fry quantity estimation model is used to identify the fry images to be detected, and the fry images to be detected in each fish pond grid area are divided into large-scale images and small-scale images through image segmentation. The feature extraction layer is distributed through a large-scale feature extraction unit and a small-scale feature extraction unit to extract features, and is processed through the SE attention mechanism to obtain a first large-scale feature map and a second small-scale feature map; then, the first large-scale feature map and the second small-scale feature map are fused according to the feature fusion layer to obtain a first fry fusion feature map; a first fry density map is obtained according to the first fry fusion feature map, and then the first fry number is obtained. This method can effectively integrate the multi-scale features of the fry images to be detected through the fry quantity estimation model, effectively improve the recognition ability of the fry images to be detected, and thus can effectively improve the accuracy of the fry quantity estimation.

[0032] Further, the second number of fry is: ; in, Indicates the second number of fry; Indicates the number of first fry; Indicates the influence weight of light intensity; Indicates the average light intensity in the fish pond; Indicates the light intensity reference value.

[0033] Further, the average light intensity of the fish pond is obtained according to the second light intensity and the number of grid areas of the fish pond; The average light intensity of the fish pond is: ; in, Indicates the average light intensity in the fish pond; Indicates the number of fish pond grid areas; represents the second light intensity; This embodiment further obtains the second number of fry by analyzing the light intensity of each fish pond grid area and the first number of fry. Through effective correction, the influence of light intensity on the estimation of the number of fry can be avoided, thereby further improving the accuracy of the estimation of the number of fry.

[0034] In this embodiment, the fish pond is firstly divided into grids, and the first image of fry to be detected, the light intensity and the area of ​​each grid area are obtained; a fry quantity estimation model is constructed to process the image to be detected, and the grayscale co-occurrence matrix of the image of fry to be detected is obtained through image recognition, and the texture feature is obtained, and then the first turbidity coefficient is obtained, and different denoising methods are determined according to the first turbidity coefficient; the feature extraction layer is processed by a large-scale feature extraction unit, a small-scale feature extraction unit and an attention mechanism; the feature fusion layer obtains the first fry fusion feature map; the feature analysis layer generates a first fry density map using the first fry fusion feature map, obtains the first fry number, and corrects it according to the light intensity to obtain the second fry number; this method can effectively improve the accuracy of fry quantity estimation.

[0035] In order to verify the effectiveness of the method for estimating the number of fry based on the multi-scale semantic fusion density estimation network provided by the present invention, the accuracy of estimating the number of fry of different methods is compared and distributed, including method one, method two and method three; method one is the method for estimating the number of fry based on the multi-scale semantic fusion density estimation network provided by this embodiment; method two is based on method one without considering denoising; method three is based on method one without considering the influence of light intensity, and the mean accuracy of the estimation of the number of fry in pond B by different methods is given, and the specific results are shown in Table 2; Table 2 Mean accuracy of different methods for estimating the number of fry in fish pond B method Estimated mean accuracy Method 1 95% Method 2 85% Method 3 84% It can be seen from Table 2 that the effectiveness of the method for estimating the number of fry based on the multi-scale semantic fusion density estimation network provided in this embodiment can effectively improve the accuracy of estimating the number of fry.

[0036] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for estimating the number of fry based on a multi-scale semantic fusion density estimation network, characterized in that: include: S1. Grid the fish pond to obtain grid areas of each fish pond, obtain the first image of the fry to be detected, the second light intensity and the third area of ​​each fish pond grid area; S2. Construct a fish fry quantity estimation model to identify the first fish fry image to be detected, the fish fry quantity estimation model includes an input layer, a preprocessing layer, a feature extraction layer, a feature fusion layer, a feature analysis layer and an output layer; the preprocessing layer performs image recognition on the first fish fry image to be detected to obtain a first turbidity coefficient; determine different denoising methods according to the first turbidity coefficient and a preset turbidity coefficient threshold set; and perform multi-scale image segmentation on the first fish fry image to be detected to obtain multiple first large-scale fish fry detection images and second small-scale fish fry detection images; S3. The feature extraction layer extracts features of the first large-scale fry detection image and the second small-scale fry detection image through a large-scale feature extraction unit and a small-scale feature extraction unit, respectively, and processes them through the SE attention mechanism to obtain a first large-scale feature map and a second small-scale feature map respectively; the feature fusion layer is used to fuse the first large-scale feature map and the second small-scale feature map to obtain a first fry fusion feature map; S4. The feature analysis layer obtains a first fry density map according to the first fry fusion feature map; obtains a first fry number according to the first fry density map and the third area; and obtains a second fry number according to the first fry number, the second light intensity and the first turbidity coefficient.

2. The method for estimating the number of fry based on a multi-scale semantic fusion density estimation network according to claim 1, characterized in that: The preprocessing layer performs image recognition on the first fry image to be detected to obtain the first turbidity coefficient specifically including: S10. graying the first fry image to be detected, and calculating the gray level co-occurrence matrix of the first fry image to be detected; S20. Extracting texture features according to the gray level co-occurrence matrix, wherein the texture features include contrast, uniformity and energy; S30. Obtain a first turbidity coefficient according to the texture feature.

3. The method for estimating the number of fry based on a multi-scale semantic fusion density estimation network according to claim 1, characterized in that: Determining different denoising methods according to the first turbidity coefficient and the preset turbidity coefficient threshold set specifically includes: S100. Determine the turbidity level of the fish pond grid area according to the first turbidity coefficient and a preset turbidity coefficient threshold set, wherein the turbidity level includes light turbidity, moderate turbidity and heavy turbidity; S200. De-noise the slightly turbid fish pond grid area by Gaussian filtering; de-noise the moderately turbid fish pond grid area by bilateral filtering; de-noise the heavily turbid fish pond grid area by deep learning denoising.

4. The method for estimating the number of fry based on a multi-scale semantic fusion density estimation network according to claim 3, characterized in that: The preset turbidity coefficient threshold set includes a first turbidity coefficient threshold and a second turbidity coefficient threshold; the specific acquisition process of the turbidity level is: ； in, Indicates turbidity level; Indicates mild turbidity; represents the first turbidity coefficient; represents the first turbidity coefficient threshold; represents the second turbidity coefficient threshold; Indicates moderate turbidity; Indicates severe turbidity.

5. The method for estimating the number of fry based on a multi-scale semantic fusion density estimation network according to claim 1, characterized in that: The fry quantity estimation model comprises an input layer, a preprocessing layer, a feature extraction layer, a feature fusion layer, a feature analysis layer and an output layer; The preprocessing layer performs image recognition on the first fry image to be detected to obtain a first turbidity coefficient, and determines different denoising methods according to the first turbidity coefficient; and performs multi-scale image segmentation on the first fry image to be detected through U-Net to obtain a plurality of first large-scale fry detection images and second small-scale fry detection images; The feature extraction layer extracts features of the first large-scale fish fry detection image and the second small-scale fish fry detection image through a large-scale feature extraction unit and a small-scale feature extraction unit, respectively, and processes them through an SE attention mechanism to obtain a first large-scale feature map and a second small-scale feature map, respectively; The feature fusion layer is used to fuse the first large-scale feature map and the second small-scale feature map to obtain a first fry fusion feature map; The feature analysis layer performs channel dimension reduction on the first fry fusion feature map through the first convolution kernel, and sequentially passes through the second convolution kernel and the third convolution kernel to obtain a first fry density map; obtains a first fry number according to the first fry density map and the third area; and obtains a second fry number according to the first fry number, the second light intensity and the first turbidity coefficient; The output layer is used to output the second number of fry.

6. The method for estimating the number of fry based on a multi-scale semantic fusion density estimation network according to claim 5, characterized in that: The first number of fry is: ； in, Indicates the number of first fry; represents the third area; Indicates the number of fish pond grid areas; represents the total number of pixels in the first fry density map; Represents the sum of pixel values ​​of the first fry density map.

7. The method for estimating the number of fry based on a multi-scale semantic fusion density estimation network according to claim 1, characterized in that: The second number of fry is: ； in, Indicates the second number of fry; Indicates the number of first fry; Indicates the influence weight of light intensity; Indicates the average light intensity in the fish pond; Indicates the light intensity reference value.

8. The method for estimating the number of fry based on a multi-scale semantic fusion density estimation network according to claim 7, characterized in that: The average light intensity of the fish pond is obtained according to the second light intensity and the number of grid areas of the fish pond.

Images