Study On Residue Recognition And Counting In Circulating Aquaculture System

Feeding is an important part of aquaculture. Not only can it improve economic benefits, but also it is the key of realizing water quality regulation.With the rapid development of intensive cultivation and the increase of labor costs, the traditional methods of artificial feeding already cannot meet the needs of intensive cultivation any more. So automated feeding systems attract vast attention and applied in production widely. At present, the domestic auto-feeding system mainly establishes feeding model based on the farming experience of fish growth for realizing the control of feeding bait. Because of the fish, water environment, feed and management etc, the fish may reduce the food intake, and the breeding pool will generate surplus food. In this case, the degree of automation of feeding system will be influenced to some extent for the lack of feedback information from the fishery. As sinking to the pond bottom, the residual baits will be filtered out through the waste collectors, and it can result in the waste of the bait. In this paper, we apply the computer vision technology to the detection of residual bait concentration, and design a bait counting method based on image recognition. The method provides a theoretic guide for real-time monitoring of bait concentrations in ponds.Firstly, the bait video image was captured at the water intake of collector.Gray-scale transformation, image binaryzation and edge detection were used to preprocess the residual bait video and obtain the outlines of the bait image. We had searched and marked bait image and keep rough count of the bait by setting the threshold of the area of the residual bait outline.Then, through the composition analysis, we knew that there were a large amount of feces and other impurities, and the main component was feces. The features(AverPixel, Peri2 Area, Conv2 Area, Skeletons, Contrast and IDM) were extracted by analyzing the difference between the bait and feces in color, shape and texture. We realized the recognition of residual bait image by using the support vector machine algorithm and the modified decision tree algorithm.Finally, through analyzing the recognition performance of the two algorithms and the decision tree algorithms was simpler, easier to transplant and better in real-time than SVM, we chose the bait recognition model that the decision tree algorithm generate to search and mark bait image, and achieve accurate counting of the bait in the video.