Real-time marine target recognition method of underwater inspection robot based on computer vision

In the complex marine environment, target recognition is difficult, and real-time detection has a slow speed. this article, method combining underwater generative adversarial network improved YOLOv4 proposed, which named M-YOLOv4. Firstly, images collected by inspection robot are enhanced using algorithm to obtain training datasets. Secondly, combines feature extraction of MoblieNetv3 for lightweight processing, reduces model size, number calculations parameters. Then, change size spatial pyramid pooling module kernel, can enlarge receptive field integrate characteristics different fields. Finally, processed datasets transferred M-YOLOv4 training, trained transplanted Jetson Nano hardware device detection. The results experiments show that mean average precision value 90.77%, 2.02% higher than unimproved one. frame per second with MobileNetv3 27, an increase 12 compared YOLOv4. perform accurate multi-targets on embedded devices.