A dual attention network based on efficientNet-B2 for short-term fish school feeding behavior analysis in aquaculture

Abstract

Fish school feeding behavior analysis based on images can provide important information for aquaculture managers to make effective feeding decision. However, it is a challenging task due to intra-class variation, cross-occlusion, and unbalanced image categories in real high-density industrial farming. At present, most of the existing works on fish school feeding behavior are limited because they seem to ignored the spatial relationship between the region of interest in fish feeding images. To address this research gap, we propose a dual attention network with Efficientnet-B2 for fine-grained short-term feeding behavior analysis of fish school. The algorithm includes EfficientNet-B2 network and two parallel attention modules, which focus on the feature extraction of the feeding region. In addition, several training strategies, such as mish activation function, ranger optimizer, label smoothing, and cosine annealing, are employed to improve the algorithm performance. Especially, label smoothing technique is used to address the problem of image class imbalance. To evaluate the effectiveness of our method, performance of proposed algorithm is analyzed on fish school feeding behavior dataset and it is also compared with benchmark Convolutional Neural Networks (CNNs) including AlexNet, VGG, Inception, ResNet, Densenet, SENet, and MobileNet. Comprehensive experimental results show that proposed algorithm achieves very good results in terms of the accuracy (the test accuracy is 89.56% on datasets), precision, parameters and floating point operations per second (FLOPS), compared with the benchmark classification algorithm. Therefore, we proposed method can be integrated into aquacultual vision system to guide farmers to plan their feeding strategy.