Accurate machine vision identification of GCHD symptom using a self-attention-based CNN model with adaptive fish separation

Grass carp hemorrhagic disease (GCHD) caused by grass carp reovirus (GCRV) is one of the most serious transmissible diseases threatening the freshwater fish aquaculture, which asks for novel efficient and costeffective surveillance techniques. Machine vision could provide an effective inspection allowing for the early warning of the GCHD epidemic spreading. But the detection accuracy using current deep learning algorithms are not satisfactory by testing the experimental datasets. In this study, we collected datasets of model animal Chinese rare minnow (Gobiocypris rarus) with GCHD symptoms and developed a self-attention-based convolutional neural network (SA-CNN) deep learning model. The SA-CNN model can accurately detect Chinese rare minnow with GCHD symptoms. After the pre-processing of the extracted images using improved k-means and enhanced watershed algorithms, the dense or overlapping fish in the populations can be adaptively separated which improves the recognizing precision. The SA-CNN model achieves an accuracy of 99.1 %, a recall of 97.7 %, and F1 score of 96.6 % with optimized loss function for the acquired datasets of fish with GCHD symptoms. The precision is higher than the current state-of-art you only look once (YOLO) series networks ranging from 4.9 % to 25.6 % with respect to different model series. The identification time for each image takes only 95ms demonstrating a good efficiency. The self-attention module brought an increase of 3.8 % in accuracy, 3.0 % in recall, as demonstrated by the ablation experiment. Thus, this optimized SA-CNN model provides an effective approach for the indispensable broad application of GCHD surveillance programs containing GCRV transmission, laying a foundation for the intelligent, healthy and precision aquaculture.