Abstract: In order to overcome the problems of underwater videos, such as low quality, blurring and even unrecognizability, using the computer vision technology for fast detection of underwater fish targets, an underwater video object detection method was proposed based on background removal methods. An object detection framework for underwater fish was designed, using the partial least squares (PLS) classifier for object detection. Input video sequences were collected from underwater fish data sets, and individual frames were extracted. After the format conversion of RGB to HSI and median filter denoising pretreatment, using the GMG background removal process, the texture and the characteristic of the gray scale coefficient were extracted based on local binary (LBP) pattern. At last, with the above extracted characteristics, the object detection of underwater fish in the daytime and night was realized using the PLS classifier. The results show that the method can achieve the object detection accuracy of 96.89% using the underwater fish video datasets, which improves the detection efficiency of underwater fish and reduces the labor cost. It can also provide some guidance for the monitoring, protection and sustainable development of underwater fish and other biological resources. 

Key words:  , partial least squares, background removal, fish, object detection, dynamic video sequence