Development and application of a deep learning algorithm-based growth phenotypes measurement system of the Pacific white shrimp (Litopenaeus vannamei)

To address the low efficiency and high error rates associated with manual measurement of growth phenotypes in the Pacific white shrimp (Litopenaeus vannamei), this study developed a dedicated image acquisition box capable of capturing standardized, high-quality side-view images of the shrimp. Utilizing this system, a High-Resolution Network (HRNet) model was employed to identify nine key feature points of the shrimp, enabling the measurement of traits such as body length. Additionally, a Mask Region Convolutional Neural Network (Mask R-CNN) model was utilized for shrimp contour segmentation to calculate body surface area. Regression models incorporating body length and body surface area were subsequently developed to predict body weight. An integrated image processing and data management software was also developed to establish a precise measurement system for the growth phenotypes of L. vannamei. The study found that the HRNet model achieved recognition rates exceeding 98% for all nine feature points, with rates exceeding 99% for seven points. The true values of body length and abdominal segment length were measured using two methods: manual measurement with a ruler and measurement from manually tagged feature points in the images. The predictive accuracy of body length and abdominal segment length was calculated to be 0.91–0.97 and 0.91–0.93, respectively, with average relative errors of 1.39%-4.63% and 2.46%-4.59%. Evaluation against manually segmented shrimp body contours showed that the Mask R-CNN model predicted body surface area with an accuracy of 0.98 and an average relative error of 1.73%. Regression models incorporating variables such as body length, body surface area, and gender were developed to predict body weight, achieving accuracies above 0.94, with the model incorporating both body length and body surface area achieving the highest prediction accuracy (0.97). These results demonstrate that computer vision technology combined with deep learning algorithms can accurately measure growth phenotypes, such as body length and body surface area, and predict body weight L. vannamei. This study provides an efficient tool for the accurate and rapid measurement of growth phenotypes in L. vannamei.