Semantic Segmentation of Freshwater Fish Body Based on Generative Adversarial Network

Abstract

The identification and precise positioning of freshwater fish heads, belly, and fins are the prerequisites for the robot to realize the key technology of fast grasping, precise cutting, and improving operation efficiency of freshwater fish. In order to solve the problem that the accuracy of identification and segmentation of freshwater fish is reduced due to imbalanced samples, poor imaging quality at the slaughter site, and small parts of the fish body. The paper proposes an algorithm for semantic segmentation of freshwater fish head, belly, and fin based on Generative Adversarial Networks (GAN). Add additional conditional features to the generator and discriminator of the image generation model to GAN theory, clearly instruct the model to generate freshwater fish information, improve image imaging quality, and expand freshwater fish images reasonably and efficiently; then input the data set into deeplabv3+semantic segmentation In the model, compared with the traditional sample expansion method, the GAN-based sample generation method better fits the semantic segmentation network, and the ratio of each part of the two groups of freshwater fish of the Mean Intersection over Union (MIOU) has reached 95%. The experimental results show that it is of great significance to apply image processing, deep learning to image optimization, expand sample data sets, and establish semantic segmentation of various parts of freshwater fish.