Automatic fish counting method using image density grading and local regression

Abstract Fish counting is a fundamental operation for biomass estimation in aquaculture. It is also vital for achieving reasonable feeding and improving breeding efficiency. Existing fish counting methods generally require each fish to be detected; in this regard, fish images with complex adhesions are difficult to process. Therefore, an automatic fish counting method using image density grading and local regression is proposed. Fish-connected areas are segmented from fish top-view images previously collected through image processing technologies. Next, four types of image features from each fish-connected area are extracted. Fish counting is realized by image density grading and local regression. Fish images are divided into several connected area sub-images, and density grading is performed on each sub-image to solve the imbalance of the fish-connected area sub-image dataset, thereby rendering more accurate and stable fish counting. The proposed method was tested on a real fish dataset, achieving a mean absolute error of 0.2985, root mean square error of 0.6105, and a coefficient of determination of 0.9607. Furthermore, compared with current typical fish counting methods, our method performed better for all evaluation metrics. Experimental results evidence the accurate fish counting accomplished by using the proposed method.