Fish Object Detection Based on Spatial Bias Pyramid Pooling: Improved BIAS-YOLO

Object detection, as an important research direction of artificial intelligence technology, has a wide scope of application in several fields. At the same time, object detection technology shows a trend of multidisciplinary integration, which frees related researchers and some laborers from repetitive labor, thus improving the efficiency and accuracy of their work. Fishery resources play a pivotal role as an important agricultural resource, and the realization of different kinds of fish detection is of great help to the rational use and protection of fishery resources. However, there are some problems in the application of object detection technology to this field, such as the lack of datasets, and the quantity and quality of samples in the data set cannot satisfy the increasingly large neural network models. In this paper, we created a dataset dedicated to fish target detection through various channels such as photography and the web. The dataset contains 16 categories and 1038 images, including common fish (e.g., carp, hairtail, etc.) and also some precious fish such as grouper. We improved the quality of the images to enable the network to learn more accurate information about the characteristics of different fish. Meanwhile, we propose BIAS-YOLO, which improves the original SPP (Spatial Pyramid Pooling) module based on the original YOLOv5 network, and introduced the idea of spatial bias, which can encode the global information and thus improve the network to make full use of the whole image information. On our proposed dataset, our network works better compared to the baseline.