MDFNet: Integrating multiscale features and dynamic attention for underwater fish counting

Counting fish is crucial for aquaculture and ecological monitoring. However, fish counting is often affected by issues such as fish overlap, varying body sizes, and complex background noise. To enhance fish counting accuracy in diverse aquatic environments, this study proposes a convolutional neural network-based accurate fish counting framework. The front-end network uses the first 10 layers of VGG 16 to extract the main feature information. First, the framework uses a multiscale fusion module to extract diverse features from the image with strip convolutions of three different scales. The framework then applies spatial attention fusion to adapt to various scale variations. Second, to address the occlusion problem in counting, a variable transformer attention module is designed to optimize the attention to the occluded area through dynamic feature interactions between each local area. Finally, a focal gated convolution module is designed to enhance the performance of the framework. Through horizontal and vertical pooling processing, the refocused convolution block and the gating mechanism are used to dynamically adjust the weighted feature map, thereby further controlling the response strength of each area in the feature map, suppressing background noise, and strengthening the fish area. Experimental results demonstrate that the model is highly effective on the common Carp Counting Dataset and the Dense Grass Carp Counting Dataset. The model offering significant potential for improving environmental monitoring and supporting sustainable fisheries practices in various aquatic ecosystems. • A multiscale feature fusion module is proposed to address the scale change problem. • A variable transformer attention module is proposed solving occlusion for fish overlap. • Noise and interference is effectively suppressed by a focal gated convolution module. • An accurate method for fish counting is designed.