An efficient fish counting method with adaptive global perception and multi-scale feature perception

Under aquaculture conditions, efficient and accurate fish counting is crucial for fishery management and ecological protection. However, existing counting methods struggle to address challenges, such as fish overlap and scale variations in the presence of complex background noise. Therefore, this study proposes a CNN-based accurate fish-counting framework. The front-end network uses the first ten layers of VGG 16 to extract the main feature information. The framework first uses a multi-scale feature perception module and four parallel-dilated convolutional networks. Each dilated convolution uses a dilated convolution with a different dilation rate to extract diverse features from the image and adapts to scale changes. Second, to reduce the occlusion problem in counting, an adaptive global perception module was designed to optimize the focus on occluded areas through information interactions between channel features. Finally, an edge excitation module was designed to reweight the features in the channel through a parallel structure using two convolutional approaches, thereby enhancing the edge feature extraction. This module addresses the issue of neglecting edge pixels, improving the model’s ability to process edge features, and reducing interference from complex background noise. Experimentally, the MAE and RMSE of the model on the carp count dataset (CCD) were 2.64 and 3.58, respectively. The average counting accuracy was 94.76%. For the dense grass carp counting dataset (DGCD), the average counting accuracy of model was 96.96%. The model performed well in terms of counting accuracy and showed good stability. Overall, this study provides strong technical support for aquaculture management and ecological protection.