Fish Bait Counting Method Based on Improved Scale Aggregation Net in Near-Infrared Images

In aquaculture, accurate counting of fish residual bait plays an important role in realizing the fine management of intelligent feeding, saving cost, reducing pollution caused by bait excess, and evaluating the feeding behavior of fish. At present, the primary methods for bait counting include acoustic-based and machine vision-based approaches, while acoustic-based technology has achieved good results, but it is susceptible to the influence of noise, and the cost of the hydrophone project used to collect the acoustic model is high, so it is not economically efficient for actual production; machine vision-based technology is more economical and mature, but the current machine vision-based bait counting method is mainly oriented to experiments with sufficient light conditions. The current machine vision-based bait counting method is mainly oriented to the experimental environment with sufficient light conditions to collect data, while some fish culture environments do not have sufficient light conditions, such as catfish culture is more suitable for breeding in gray conditions, for the lack of light can be solved by using near-infrared (NIR) technology to collect data, but the resolution of the images collected by NIR technology is relatively low, and the information in NIR images is less than the information in color images and there is electrical noise that affects the imaging effect. Therefore, this paper proposes an improved adaptive scale aggregation network based on SANet(Scale Aggregation Network) to realize the task of accurate counting of residual baits in blurred images under NIR conditions. The mean error of the Adaptive Scale Aggregation Network (ASANet) is 10.75, which is a 32% reduction in the mean absolute error of the Adaptive Scale Aggregation Network (ASANet) compared to the original SANet network. In summary, the bait particle counting model proposed in this paper has high counting accuracy and is more stable, which is conducive to providing theoretical support for aquaculture intelligent feeding research.