An assessment method for the severity of shrimp red-leg disease based on E-DarkNet53 and multichannel feature fusion deep learning network

Red-leg disease is a common disease in shrimp farming and is characterized by high mortality rates and challenges in detecting and assessing the severity of infection. Delayed disease detection can result in large-scale outbreaks. To address the challenges of detection and the high false-positive rates associated with red-leg disease in intelligent shrimp farming, this research proposes a multibranch, multichannel feature information fusion approach to assess the infection severity of shrimp red-leg disease. First, a weight enhancement mechanism is proposed to integrate key feature channels into DarkNet53, and an idea is proposed that unifies the activation functions of SENet and DarkNet53 by using Leaky ReLU. Second, this research proposes a multibranch, multichannel feature information fusion classification network called E-DarkNet53, which integrates an enhanced SENet mechanism into the branch module of the DarkNet block. Finally, the dropout mechanism is embedded into DarkNet53, along with an additional connection layer to map the hidden-layer feature space to the sample labeling space which mitigates the risk of overfitting. E-DarkNet53 is compared with 14 mainstream classification networks to verify its effectiveness. Experimental results show that E-DarkNet53 achieved a test accuracy of 92.87%, outperforming the highest accuracy model among the 14 mainstream networks by 1.43%. This research provides an effective approach for the intelligent detection of the infection severity of shrimp red-leg disease and can be extended to other biological disease applications.