Fish key feature point detection and sign identification based on deep learning

In order to ensure real-time monitoring of growth conditions and scientific breeding management in the process of fish farming, it is necessary to realize efficient and automatic fish sign recognition. Based on this, a fish sign recognition method based on deep learning key feature point detection model combined with binocular vision is proposed. Based on the preprocessed monocular vision data set, the high-resolution network model integrated into the pyramid segmentation attention is trained to obtain the fish key feature point detection model. On this basis, the binocular vision image can be rapidly detected, recognized and matched with each feature point, and the real coordinates of each feature point and corresponding physical parameters can be calculated according to the internal parameters of the binocular vision system The test results show that the PCK value of the established key feature point detection model for each feature point is greater than 0.85, and the relative error of the identified sign parameters is less than 10%, which can provide support for the rapid identification of fish signs and effectively help the scientific and intelligent development of fish farming.