Improved YOLO-InceptionV3 network for shrimp detection in recirculation aquaculture system

Shrimp detection on the feeding platform constitutes a crucial component in recirculation aquaculture system. But automated detection faces several difficulties such as complicated background of shrimps feeding platform, overlapping of high-density shrimps, variable postures of shrimps, which reduce the accuracy of detection. In this paper, InceptionV3 modules are designed and You Only Look Once-InceptionV3 (YOLO-InceptionV3) is proposed. Besides, we introduce Efficient Intersection over Union (EIoU) to YOLO-InceptionV3 to further enhance localization accuracy. The YOLO-InceptionV3 improves shrimps feature extraction and enhances shrimp localization capabilities, mitigating the problems raised by complex background, overlapping shrimps and variable postures. We compare YOLO-InceptionV3 with other methods, as well as YOLO models with different backbone architectures and activation loss functions. Experimental results demonstrate YOLO-InceptionV3 attains satisfactory performance, achieving 0.9883 of precision, 0.9713 of recall, 0.9906 of mAP@0.5, 0.7559 of mAP@0.5:0.95 with a compact model structure. These results confirm that YOLO-InceptionV3 enables precise shrimp detection and can be deployed easily in the complex scenario of recirculation aquaculture system.