Fish mass estimation method based on adaptive parameter tuning and disparity map restoration under binocular vision

Abstract

Accurately estimating fish mass is crucial for evaluating fish growth status, enabling precise feeding, and improving aquaculture efficiency. In current aquaculture, invasive measurements may cause certain damage to fish and affect their healthy growth, non-invasive measurements mostly use stereo cameras to capture fish images and extract two-dimensional and three-dimensional features for mass estimation. The calculation of three-dimensional fish features relies on the acquisition of accurate disparity maps. Most current disparity maps are obtained through manual tuning of algorithmic parameters, which not only increases labor and time costs but also introduces a degree of randomness. To address the above problems, a fish mass estimation method based on Adaptive Parameter Tuning-Disparity Map Restoration (APT-DMR) and Multiple Linear Regression (MLR) under binocular vision technology is proposed. Firstly, fish images are obtained using a binocular camera, followed by camera calibration and image correction. Secondly, image processing technologies are used to segment the corrected image to obtain the fish target, and the two-dimensional features of the fish target are extracted. On this basis, the method based on APT-DMR is adopted to obtain the fish disparity map, extract the corresponding key matching points of the left and right images of the fish, and calculate the coordinates of the three-dimensional spatial feature points using the triangular transformation principle, achieving the extraction of the three-dimensional features of the fish target. Finally, fish mass is predicted using the MLR method. Based on the binocular vision technology, the APT-DMR method is employed to obtain the fish disparity map, which realizes the extraction and calculation of the three-dimensional features of the fish. The proposed method effectively addresses the problem that the traditional algorithm needs to constantly tune the parameters to obtain an accurate disparity map. Additionally, a new feature, the fish depth ratio, is introduced to enrich the model representation, and finally, the fish mass is successfully predicted. In addition to saving time and labor costs, the proposed method also effectively eliminates the stress and potential damage to the fish caused by invasive mass measurements. The crucian carp were taken as the experimental object and the proposed method was tested on the real dataset. The results show that the mean absolute error (MAE) is 0.0061, the root mean square error (RMSE) is 0.0084, and the coefficient of determination (R²) is 0.9338. Compared with Polynomial Regression (PR)-Weight, Decision Tree Regression (DTR)-Weight, Random Forest Regression (RFR)-Weight, Back-Propagation Neural Network (BPNN)-Weight, and Support Vector Regression (SVR)-Weight mass estimation methods, the performance of each evaluation metric of the proposed method has been greatly improved, predicting the fish mass more accurately.