DF-DETR: Dead fish-detection transformer in recirculating aquaculture system

Abstract

In aquaculture, real-time and rapid detection of dead fish is important for early risk warning and improving aquaculture efficiency. However, the complex actual environment and uncontrollable fish movement have brought great challenges to the detection of dead fish. Therefore, this paper proposes a high-precision and lightweight dead fish-detection transformer (DF-DETR) based on machine vision and original RT-DETR (real-time detection transformer). The specific implementation is as follows: Firstly, the backbone of the original RT-DETR was replaced by the RepNCSPELAN module which extracts multi-scale features. This not only improves the model’s ability to detect targets of different sizes but also reduces the amount of model parameters. Secondly, the AIFI in the RT-DETR was improved to CascadedGroupAttention (CGA). By changing the original feature fusion method, different levels of features are grouped and attention mechanism is added, so as to capture more target features. Finally, the CCFM_CSP module was constructed to fuse important features using parallel dilated convolution with different expansion rates, which improves the detection accuracy. The experimental results show that the mAP@.5 of the proposed dead fish detection model DF-DETR can reach 96.6%, and the parameter amount is reduced by 27% compared with the original RT-DETR. In summary, the proposed DF-DETR model realizes real-time and high-precision dead fish detection, which can provide effective technical support for the development of intelligent inspection robots.