Based On Computer Vision The Counting System Of Puffer Fish

In the process of aquaculture,the detection and counting of fish in a pond is a necessary condition for estimating fish density,length and weight.At present,there are still using the method of manual counting in factories,which is inefficient and time-consuming.In recent years,computer vision technology is one of the hot spots.The typical algorithm YOLOv3 has been widely applied in real life,but there are some shortcomings in the enumeration of cultured puffer fishes.Scientists research a non-contact and automatic counting technology based on computer vision,which will greatly improve the intelligence and automation of the cultivation of puffer fishes.Therefore,this paper will develop a computer vision based counting system for puffer fishes based on the characteristics of fish data set in the ponds.This paper is summarized as follows:1)Research on the augmentation method of fish data set in aquaculture ponds.In view of the problems of low quality,labeling difficulty and obscuration of fishes in real culture environment,so it is proposed an integrated method of fish data set augmentation in culture ponds.This method can effectively solve the over fitting problem in the process of model learning.According to the characteristics of darkening,blurring and occlusion in the fish groups,this method proposed three data enhancement methods: brightness enhancement,blurring enhancement and partial occlusion enhancement,which further enriched the number of images.Experiments show that three data augmentation methods in this study can effectively improve the recognition and detection effect.2)Research on an improved detection and counting algorithm based on yolov3 model.In view of the fact that the target size in the cultured fish images are relatively uniform and little difference.So this paper is proposed a recognition and counting method based on improved yolov3.In this method,k-means algorithm is used to calculate the size of the target,and obtain the distribution characteristics of the target size.Based on the statistical results,the yolov3 model is optimized to a simplified model with only medium scale detection and small scale detection.On the one hand,this method reduced the complexity of the model and improved the recognition effect of the model.And experiments select soft-NMS algorithm to filter the border,which effectively reduces the possibility of the target being deleted by mistake and the target being missed detection.Experiments show that in new algorithm,the counting accuracy is more than 90%,the recall rate is more than 93%,and the detection speed is improved by 3MS,which is greatly improved compared with the traditional model yolov3.