Counting method for cultured fishes based on multi-modules and attention mechanism

Abstract

Fish counting is of great significance for aquaculture companies to formulate feeding strategies and management plans. Preceding computer vision methods are difficult to solve the counting problem when the fish body is seriously overlapped and the shape changes greatly. In this research, a deep learning network model based on multi-modules and attention mechanism (MAN) is proposed to realize the counting of cultured fish which consists of feature extraction module, attention module, and density map estimation module. Among them, the feature extraction module is composed of three parallel convolutional networks which is used to extract the general feature map of the image and serves as the input of the subsequent module. The trunk convolutional network and the attention network are connected in parallel to form the attention module which is more precisely identify critical information of image in dense counting process. Thereinto, the attention network integrated a residual block without batch normalization and a convolutional layer. The distribution and the number of fishes in the image is represented by density map estimation module. In order to improve the counting accuracy and retain more details of the image, the sum of squared error (SSE) and structural similarity index (SSIM) are applied to form the loss function and train the model. In order to explore the generalization ability of the model, data from two tanks were used to validate the model. The experimental results for MAN showed that the counting accuracy is about 97.12%, and the deviation is 3.67. Compared with the MCNN and CNNs, the accuracy of MAN has been improved by about 1.95% and 2.76%. At the same time, the deviation of MAN has also been reduced by 30.23% and 41.09%, respectively. For the images of another tank, the counting accuracy of MAN is 94.69%, while the MCNN and CNNs cannot meet the application requirements of fish counting. Furthermore, the accuracy for MAN between the different distribution intervals of the number of fishes is fluctuating less. On the whole, the model not only has high counting accuracy, but also has strong stability and generalization performance. The method that based on multi-modules and attention mechanism can promote the exploration of the technology related to dense fish counting, and provide a feasible solution for the application of this technology in practical engineering.