A non-destructive method for predicting the flesh yield of largemouth bass
By constructing a total meat quality prediction model and using the YOLO algorithm to automatically identify key points, the problem of slaughtering required for calculating the meat yield of largemouth bass was solved, achieving non-destructive, rapid, and accurate prediction, and supporting efficient screening for largemouth bass breeding.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING INSTITUTE OF FISHERY SCIENCES (NANJING AQUATIC TECHNOLOGY PROMOTION STATION NANJING AQUATIC ANIMAL DISEASE PREVENTION & CONTROL CENTER)
- Filing Date
- 2026-03-11
- Publication Date
- 2026-06-19
AI Technical Summary
Current methods for calculating the meat yield of largemouth bass require slaughtering the fish, which is a cumbersome and time-consuming process that cannot be repeatedly monitored, thus limiting its application in breeding.
By acquiring morphological parameters and quality data of largemouth bass, a total meat weight prediction model is constructed. High-resolution cameras are used to photograph live fish and extract morphological parameters. Combined with the YOLO algorithm, key points are automatically identified, and the meat yield is calculated for lossless prediction.
It achieves non-destructive, rapid, and accurate prediction of meat content, reduces reliance on operational skills, improves detection efficiency, is suitable for repeated testing in live organisms, and supports the screening of superior varieties in breeding projects.
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention relates to the field of aquatic breeding technology, specifically a non-destructive method for predicting the meat yield of largemouth bass. Background Technology
[0002] Largemouth bass, an important freshwater aquaculture species in my country, is highly favored by farmers and consumers due to its tender flesh, rich nutrition, and rapid growth. Its aquaculture scale has been expanding year by year, making it one of the leading species in the aquaculture industry. Meat yield is a core economic trait for evaluating the aquaculture value and breed quality of largemouth bass, directly affecting aquaculture efficiency and market acceptance. Therefore, meat yield is a crucial screening indicator that must be focused on during the breeding of superior largemouth bass.
[0003] Current methods for calculating the meat yield of largemouth bass require slaughtering the fish and separating the muscle, then dividing the muscle mass by the total fish mass. This method has significant drawbacks: firstly, the calculation process is cumbersome and time-consuming, relying heavily on the operator's anatomical skills; secondly, it damages the sample fish, making it impossible to preserve the live specimens for multiple follow-up measurements during breeding, and also hindering breeding stock selection, severely limiting its application in selective breeding. Therefore, a non-destructive, rapid, and accurate method for predicting meat yield is urgently needed to meet the requirements of largemouth bass breeding projects. Summary of the Invention
[0004] The purpose of this invention is to provide a non-destructive method for predicting the meat yield of largemouth bass, in order to solve the problems raised in the prior art.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a non-destructive prediction method for the meat yield of largemouth bass, comprising the following steps: S1, acquiring a batch of morphological parameters and mass data of largemouth bass, and constructing derived features based on the morphological parameters; S2, training a total meat weight prediction model based on the morphological parameters, mass data, and derived features obtained in step S1, and further evaluating the performance of the total meat weight prediction model; S3, placing the live largemouth bass to be tested on a collection platform to weigh the total mass of the fish, and using a high-resolution industrial camera to perform standardized photography of the live largemouth bass and automatically extracting the morphological parameters of the largemouth bass, constructing corresponding derived features; S4, inputting the morphological parameters and derived features obtained in step S3 into the total meat weight prediction model, and the total meat weight prediction model calculates and outputs the predicted total meat weight value of the largemouth bass; S5, obtaining the predicted meat yield value of the largemouth bass according to the formula: predicted total meat weight / total fish mass.
[0006] As a preferred technical solution, the morphological parameters in step S1 are the total length, body length, head length, body depth, caudal peduncle depth, caudal peduncle length, eye diameter, and snout length of the largemouth bass; the quality data are the mass of the fins, scales, viscera, total meat, abdominal meat, back meat, and tail meat of the largemouth bass, accurately measured by traditional slaughtering methods; the derived characteristics are the total fish area and body area of the largemouth bass.
[0007] As a preferred technical solution, the total meat quality prediction model described in step S2 enhances diversity during training by using Bootstrap sample sampling and random feature selection.
[0008] As a preferred technical solution, the formula for the total meat weight prediction model in step S2 is: Where y^ is the predicted total meat quality, f t (x) represents the predicted output of the t-th decision tree, where x represents the input features, i.e., morphological parameters and derived features; the number of decision trees t ranges from 100 to 300.
[0009] As a preferred technical solution, the total meat quality prediction model uses the mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), Pearson correlation coefficient (Pearson r), and p value as evaluation indicators.
[0010] As a preferred technical solution, the standardized shooting in step S3 refers to shooting side and top view images of the largemouth bass.
[0011] As a preferred technical solution, the standardized side and top view images are used to identify key points using the YOLO algorithm model, and the morphological parameters of the largemouth bass are automatically extracted through image processing algorithms, namely the total length, body length, head length, body depth, caudal peduncle depth, caudal peduncle length, eye diameter, and snout length of the largemouth bass, with an extraction error of ≤ 0.5 mm.
[0012] As a preferred technical solution, the key identification points include 14 points, which are specifically as follows: Point 1: the point marked at the very front of the upper lip of the fish head; Point 2: the point where the lateral line scale, head scale, and gill cover intersect; Point 3: the point at the highest point of the upper fish width; Point 4: the point at the bottom corresponding to the vertical line of the fish body at the highest point of the upper fish width; Point 5: the point at the narrowest point at the upper end of the caudal peduncle; Point 6: the point at the base of the anal fin; Point 7: the point at the bottom of the narrowest point at the lower end of the caudal peduncle, perpendicular to the head-tail axis; Point 8: the point midway between the central axis of the fish body and the concave position of the fish body and tail; Point 9: the point at the very top of the tail; Point 10: the point directly above Point 6 where it intersects with the lateral line scale; Point 11: the point at the upper edge of the eye socket; Point 12: the point at the lower edge of the eye socket; Point 13: the point at the anterior edge of the eye socket; Point 14: the point at the intersection of the upper and lower lips.
[0013] As a preferred technical solution, the process of automatically extracting morphological parameters using the YOLO algorithm model includes: normalizing the size of the acquired side-view and top-view images; extracting multi-scale depth features of the images using a multi-layer convolutional neural network; outputting the confidence heatmap and coordinate offset of each key point through a regression branch; then removing redundant boxes and key recognition points with low confidence; and accurately locating the coordinates of the key recognition points. During the training of the YOLO algorithm model, manually labeled large black bass are used, covering different orientations and poses. YOLOv11-Pose is used as the basic network architecture, and pre-trained weights loaded on the COCO large-scale general dataset are used as initialization parameters to accelerate model convergence. The training runs for 300 epochs, and a cosine annealing strategy is introduced to dynamically adjust the learning rate.
[0014] As a preferred technical solution, the derived features are obtained by extracting the masked regions of the whole fish and the fish body in the image through the convolutional neural network in the YOLO algorithm model, calculating the number of pixels in the two masked regions respectively, and then converting them according to the pixel-to-actual area ratio to obtain the derived features, namely the whole fish area and the fish body area. The whole fish area (All region) is defined as the maximum outer contour of the largemouth bass, and the labeled range includes: head, trunk, tail, and all attached fins (dorsal fin, anal fin, caudal fin, pectoral fin, and pelvic fin). The fish body area (Body region) is defined as the main trunk of the largemouth bass, and the labeled range includes: the whole fish area (All region) excluding the head and all fins, with the labeled boundary cut along the connection between the head, fins, and body.
[0015] Compared with the prior art, the beneficial effects of the present invention are:
[0016] 1. This invention completely eliminates the destructive steps of slaughtering and removing meat in traditional methods. It only requires photographing and weighing the live fish, with each sample taking about 3 seconds to measure, which greatly improves the detection efficiency.
[0017] 2. The measurement process of this invention does not require harming the fish and the measurement time is extremely short. Therefore, it does not damage the sample fish and can be directly applied to the breeding project of largemouth bass to conduct live and repeated tracking and measurement of meat content of candidate parents or different families, and screen out high meat content superior varieties.
[0018] 3. This invention reduces the reliance on the operator's anatomical skills. Through standardized image acquisition and automated algorithm analysis, it reduces human error and makes the results more consistent and comparable. Attached Figure Description
[0019] Figure 1 This is a diagram showing the key identification points in this invention; Figure 2 This is a diagram showing the morphological parameters of the fish in this invention. Figure 3This is a mask diagram showing the total area of the fish and the area of the fish body in this invention; Figure 4 This is a graph showing the performance test results of the total meat quality prediction model in this invention. Detailed Implementation
[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0021] Example
[0022] This embodiment provides a non-destructive method for predicting the meat yield of largemouth bass, including the following steps:
[0023] S1. 485 largemouth bass samples were selected, and the morphological parameters of this batch of largemouth bass samples were measured using a data acquisition platform, namely total length AL, body length BL, head length CL, body depth CH, caudal peduncle depth TH, caudal peduncle length TL, eye diameter EL, and snout length ML. Then, the quality data of fins, scales, internal organs, total meat, abdominal meat, back meat, and tail meat were accurately measured using traditional slaughtering methods. Finally, the derived characteristics were calculated based on the morphological parameters, that is, the total fish area all and body area of the largemouth bass sample were calculated. Some data are shown in Table 1 and Table 2.
[0024] Table 1 shows the morphological parameters of artificially measured largemouth bass samples (partial data), in mm; Table 2 shows the mass data and derived characteristics of artificially measured largemouth bass samples (partial data), with mass in g and area in pixels.
[0025]
[0026] Table 1
[0027]
[0028] Table 2
[0029] S2. Divide the morphological parameters, quality data, and derived features obtained in step S1 into a dataset of 80% training set and 20% test set. Set a random number (random_state=42) to ensure the results are reproducible. Set the morphological parameters and derived features as inputs and the total meat quality as output. During training, the model is enhanced with two stages of randomness to increase diversity: 1. Bootstrap sample sampling: Perform T samplings with replacement in the training set for prediction training, where T is the number of decision trees set, which is between 100 and 300. In this embodiment, T=200.
[0030] 2. Random Feature Selection: When the decision tree splits at each node, it selects only the optimal splitting feature, i.e., the input morphological parameters. After training, the output total meat quality prediction model is the average of the predictions from all decision trees, expressed by the formula: , where y^ is the final predicted value, T is the number of decision trees, ft(x) is the predicted output of the t-th decision tree, and x is the input feature, namely the morphological parameters and derived features.
[0031] Based on the final output of the model, the performance of the total meat quality prediction model was evaluated. The mean absolute error (MAE), mean squared error (MSE), root mean square error (RMSE), Pearson correlation coefficient (Pearson r), and p-value were used as evaluation indicators. Figure 4 For the test results of the model, from Figure 4 The results show that: MAE = 3.8600, MSE = 26.4561, RMSE = 5.1435, R = 0.9361, and p-value = 0.0000.
[0032] MAE, MSE, and RMSE (error metrics) determine whether the model's accuracy meets business requirements. MAE = 3.8600: The average deviation of each prediction from the true value is 3.86 units. MSE = 26.4561, RMSE = 5.1435: More sensitive to large errors, reflecting the fluctuation of predicted values. The target quantity is weight, which ranged from 230 to 310 in this test, with an average value of 259.67. The deviation of 3.86 units from 259.67 is very small, indicating high model accuracy and meeting the usage requirements.
[0033] R (correlation coefficient) and p-value (significance test) are used to determine the effectiveness of the model. R represents the linear correlation between predicted and actual values, ranging from -1 to 1. The closer to 1, the better the fit. Generally, R > 0.9 indicates a strong correlation. In this model, R = 0.9361, indicating that the model's predicted trend is highly consistent with the actual situation. p-value represents the probability that this correlation is accidental. Generally, p < 0.05 indicates significance. A p-value of 0.0000 indicates that this high correlation in this test is not coincidental, and the model is statistically significant and effective. Based on these two indicators, the model has been successfully established and has a high goodness of fit.
[0034] S3. Ten live largemouth bass were placed on separate collection platforms and their total weight was measured. Standardized images of the live largemouth bass were then taken using a high-resolution industrial camera. The side and top views obtained after the standardized images were first annotated with key identification points using the YOLO image model. The key identification point features are as follows: Figure 1 As shown, the positions of the morphological parameters on the fish body are as follows: Figure 2 As shown in Table 3 below, the total mass of the fish is as follows;
[0035] Then, the morphological parameters of the largemouth bass were automatically extracted using key identification points. Image processing algorithms were used to normalize the size and pixels of the acquired side and top view images. Multi-layer convolutional neural networks were used to extract multi-scale depth features of the images, capturing the edge, texture and anatomical structure information of the fish body. The confidence heatmap and coordinate offset of each key point were output through the regression branch. Redundant boxes and key identification points with confidence scores below 0.6 were then removed to accurately locate the coordinates of the key identification points. Finally, based on the actual length of the fixed shooting, the pixel size was converted into the actual length, thereby realizing the automatic extraction of morphological parameters, namely total length, body length, head length, body height, caudal peduncle height, caudal peduncle length, eye diameter and snout length, with an extraction error ≤ 0.5 mm. The morphological parameter extraction results are shown in Table 3 below.
[0036] Then, using 14 key recognition points, the mask extracted from the largemouth bass was identified, specifically the mask for the fish body and the overall mask. The detection area is as follows: Figure 3 As shown, the number of pixels in the two mask areas is calculated to obtain the derived features of the largemouth bass, namely the total fish area and the fish body area. The results of the derived features are shown in Table 3 below.
[0037] Table 3 shows the total body mass of the largemouth bass samples in g; Table 4 shows the morphological parameters of the automatically extracted largemouth bass samples in mm; Table 5 shows the derived features of the automatically extracted largemouth bass samples, with the area in pixels.
[0038]
[0039] Table 3
[0040]
[0041] Table 4
[0042]
[0043] Table 5
[0044] S4. Input the morphological parameters and derived features obtained in step S3 into the total meat quality prediction model trained in step S2. The total meat quality prediction model calculates and outputs the total meat quality prediction value of the largemouth bass. The results are shown in Table 6 below.
[0045]
[0046] Table 6
[0047] S5. Based on the predicted meat yield value = predicted total meat mass / total fish mass, input the data obtained in step S4 to obtain the predicted meat yield value of the largemouth bass. The results are shown in Table 6 below.
[0048] Table 7
[0049] S6. Continue to use the traditional slaughtering method to accurately measure the actual meat yield of the 10 largemouth bass samples after the measurement in step S4. Compare the actual meat yield with the predicted meat yield as shown in Table 8 below.
[0050]
[0051] Table 8
[0052] As shown in Table 8, the difference between the actual and predicted meat content is very small, within the allowable error range. Therefore, the total meat quality prediction model established in this invention is highly accurate and meets the requirements. The accuracy of the results demonstrates that the YOLO image model, during image processing, accurately utilizes key identification point annotation to obtain morphological parameters and derived features. This completely eliminates the destructive step of slaughtering and harvesting meat required in traditional methods, requiring only photographing and weighing of live fish. Each sample measurement takes approximately 3 seconds, greatly improving detection efficiency. With its advantage of causing no damage to the sample fish, this method can be directly applied to largemouth bass breeding projects, allowing for live, repeated meat content tracking measurements of candidate parents or different families to screen for high-meat-content superior varieties.
[0053] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A method for non-destructive prediction of the flesh content of a largemouth bass, characterized in that, Includes the following steps: S1. Obtain a batch of morphological parameters and quality data of largemouth bass, and construct derived features based on the morphological parameters; S2. Based on the morphological parameters, quality data, and derived features obtained in step S1, a total meat quality prediction model is trained, and the performance of the total meat quality prediction model is further evaluated. S3. Place the live largemouth bass to be tested on the collection platform and weigh the total mass of the fish. Use a high-resolution industrial camera to take standardized pictures of the live largemouth bass and automatically extract the morphological parameters of the largemouth bass to construct the corresponding derived features. S4. Input the morphological parameters and derived features obtained in step S3 into the total meat quality prediction model. The total meat quality prediction model calculates and outputs the predicted total meat quality of the largemouth bass. S5. Based on the predicted meat yield value = predicted total meat mass / total fish mass, the predicted meat yield value of the largemouth bass is obtained.
2. The method of claim 1, wherein the method is a non-destructive method of predicting the flesh yield of a largemouth bass. The morphological parameters mentioned in step S1 are the total length, body length, head length, body depth, caudal peduncle depth, caudal peduncle length, eye diameter, and snout length of the largemouth bass; the mass data are the mass of the fins, scales, viscera, total meat, abdominal meat, back meat, and tail meat of the largemouth bass, accurately measured by traditional slaughtering methods; the derived characteristics are the total fish area and body area of the largemouth bass.
3. The method of claim 1, wherein the method is a non-destructive method for predicting the flesh yield of a largemouth bass. In step S2, the total meat quality prediction model enhances diversity during training by using Bootstrap sample sampling and random feature selection.
4. The method of claim 1, wherein the method is a non-destructive method for predicting the flesh yield of a large-mouth bass. The formula for the total meat weight prediction model mentioned in step S2 is: where y^ is the predicted value of total meat quality, f t (x) is the prediction output of the tth decision tree, and x is the input feature, i.e., the morphological parameters and derived features.
5. The method of claim 1, wherein the method is a non-destructive method for predicting the flesh yield of a largemouth bass. The total meat quality prediction model uses mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), Pearson correlation coefficient (Pearson r), and p-value as evaluation indicators.
6. The method of claim 1, wherein the method is a non-destructive method for predicting the flesh yield of a large-mouth bass. The standardized shooting mentioned in step S3 refers to taking side and top view images of the largemouth bass.
7. The method of claim 6, wherein the method further comprises: The standardized side and top views were captured using the YOLO algorithm model for key point recognition. The morphological parameters of the largemouth bass were automatically extracted using image processing algorithms, namely the total length, body length, head length, body depth, caudal peduncle depth, caudal peduncle length, eye diameter, and snout length, with an extraction error of ≤ 0.5 mm.
8. The method of claim 7, wherein the method further comprises: The key identification points include 14 points, and the key identification points are as follows: Point 1: The point marked at the very tip of the upper lip, where the fish's head is located; Point 2: The point where the lateral line scale, cephalic scale, and gill cover intersect; Point 3: The highest point of the upper fish width; Point 4: The point at the bottom corresponding to the highest point of the fish's width along the vertical line from the middle of the fish's body; Point 5: This is the point at the narrowest point at the upper end of the fish tail peduncle; Point 6: The point at the base of the anal fin; Point 7: The point at the bottom of the narrowest part of the fish tail peduncle, perpendicular to the head-tail axis; Point 8: The point midway between the central axis of the fish's side and the depression at the tail. Point 9: The very tip of the tail; Point 10: The point where it intersects with the lateral line scale directly above point 6; Point 11: This is the point on the upper edge of the eye socket; Point 12: This is the point at the lower edge of the eye socket; Point 13: This is the point on the anterior margin of the eye socket; Point 14: The point where the upper and lower lips meet.
9. The method of claim 8, wherein the method further comprises: The process of automatically extracting morphological parameters using the YOLO algorithm model includes: normalizing the size of the acquired side and top view images; extracting multi-scale depth features of the images using a multi-layer convolutional neural network; outputting the confidence heatmap and coordinate offset of each key point through a regression branch; then removing redundant boxes and key recognition points with low confidence; and accurately locating the coordinates of the key recognition points.
10. The method of claim 9, wherein the method further comprises: The derived features are obtained by extracting the masked regions of the whole fish and the fish body in the image through the convolutional neural network in the YOLO algorithm model, calculating the number of pixels in the two masked regions respectively, and then converting them according to the pixel-actual area ratio to obtain the derived features, namely the area of the whole fish and the area of the fish body.