Defect detection method of Chinese mitten crab based on improved EfficientVit

An improved EfficientVit-based defect detection method for Chinese mitten crabs (Eriocheir sinensis) is proposed to enhance detection efficiency and accuracy while reducing human interference. A dataset of intact and defective crab images was constructed to improve the model's adaptability. The squeeze-and-excitation (SE) module and mix of experts hidden (MoH) module were integrated into the EfficientVit model to strengthen its feature extraction ability and sensitivity to subtle differences. The ReLU activation function was replaced with the parametric rectified linear unit (PReLU) activation function to address the 'dying neuron' issue and improve the model's sensitivity to negative inputs. The improved model achieved an accuracy of 97.34%, precision of 97.35%, recall of 97.34%, and F1-score of 97.34%, showing significant improvements over the baseline model. Future work will focus on optimising model performance and exploring its adaptability and generalisation in different environments.