Study on Real-time Recognition of Underwater Live Shrimp by the Spherical Amphibious Robot Based on Deep Learning

Abstract

In this paper, spherical robots are used for the detection and identification of lobsters in aquaculture. Lobster farmers are often faced with tasks such as observation, feeding, and fishing, which are all done manually, with low efficiency and high operating costs. Therefore, this paper proposes a real-time underwater lobster detector based on Generative Adversarial Networks and Convolutional Neural Networks, implemented by a spherical amphibious robot. Firstly, the underwater lobster image dataset is established, and the improved GAN algorithm and data increment method are used for data enhancement preprocessing. Secondly, the single-shot multi-frame detector (SSD) is improved as follows, using the lightweight network MobileNetV2 as the backbone of the SSD network; in the network prediction layer, using depthwise separable convolution instead of standard convolution to accelerate inference; compressing the fully connected layer The parameters construct a lightweight model. Finally, the model is trained on the underwater lobster dataset and deployed on a spherical amphibious robot, and the changes in the loss function value during training before and after image enhancement and algorithm improvement are plotted. Two sets of experimental test results show that the model optimizes the target recognition accuracy of underwater lobsters, and the recognition accuracy reaches 90.32%. The reduced model size facilitates model deployment and is only 24MB in size. The model has good stability and high recognition accuracy in identifying lobsters in complex situations.