CAGNet: an Improved Anchor-Free Method for Shrimp Larvae Detection in Intensive Aquaculture

Guoxu Zhang,Zhencai Shen,Daoliang Li,Ping Zhong,Yingyi Chen

DOI: https://doi.org/10.1007/s10499-024-01460-0

2024-01-01

Aquaculture International

Abstract:Object detection adopting deep-learning has strongly promoted the development of intensive aquaculture. However, shrimp larvae, as an important aquatic organism, are more difficult to be detected than others. On the one hand, they have indeed small sizes, which will cause them to be easily ignored due to the background noise pollution. On the other hand, affected by environmental factors and the fact that shrimp larvae like to move fast as jumping, the images of shrimp larvae often appear blurry. In order to obtain better shrimp larvae detection performance, we propose an improved anchor-free method called CAGNet in this paper. Compared with YOLOX_s, three structures including backbone, neck, and head have been improved in the proposed method. Firstly, we ameliorate the backbone by adding a coordinate attention module to extract more location information and semantic information of shrimp larvae at different levels. Secondly, an adaptively spatial feature fusion module is introduced to the neck. It can adaptively integrate effective shrimp larvae features from different levels and suppress the interference of conflicting information arising from the background. Moreover, in the head, we use GIoU module instead of conventional IoU for more accurate bounding box regression. We conducted experiments by collecting shrimp larvae data from a real aquaculture farm. Compared with the general object detection methods and previous related research, CAGNet has achieved better performance in Precision, Recall, F1 Score, and AP@0.5:0.95. Hence, the proposed method can be effectively applied to shrimp larvae detection in intensive aquaculture.