Research On Fish Object Detection And Image Segmentation In Aquaculture Environment

China is a traditional fishing country,and a large amount of labor force needs to be invested in fishery production.In recent years,information technology has developed rapidly.How to use information technology to improve the efficiency of fishery production,realize the automation and intelligence of fishery production,and reduce labor costs has important research significance.In the process of fishery breeding,it is necessary to monitor fish regularly.Machine learning method can be used to extract the biological information of fish and realize automatic monitoring.However,in the aquaculture environment,the underwater image has different characteristics from the shore image,such as color deviation,uneven contrast and blur.This paper focuses on the target detection and segmentation method of fish image in the aquaculture environment.Through the analysis of literature,it is concluded that the current research has the following deficiencies: 1)the current research on fish target detection mainly takes the clearer underwater image as the data,and it is difficult to consider the difficulty of target detection in fuzzy underwater image.2)Many fish have the habit of cruising and foraging at night.The current research on fish target detection is mainly based on daytime data,which has not effectively solved the problem of target detection in fish images at night or under insufficient illumination.3)Under the breeding conditions,fish come in groups to look for food.Due to the mutual shielding and truncation of fish in the shooting picture,there will be interference.If the current method is used to carry out fish quantity statistics,there will be a serious problem of missing inspection.4)To monitor the growth of cultured fish,it is necessary to accurately grasp the shape of fish body.Due to the fuzzy underwater image and mixed fish targets,the current image segmentation method has the problem of low segmentation accuracy of fuzzy underwater fish image.Based on the above problems,this paper focuses on the methods of fish target detection and image segmentation,and obtains the following research results:1)Aiming at the lack of fish target detection data set in aquaculture environment,a fish data set reflecting the real-time state under aquaculture conditions is established.Most of the existing public fish image classification or segmentation data sets are images taken in the sea or after fishing,or mainly small fish.The images are relatively clear,which is not in line with the real scene in the breeding environment.In order to better carry out the work of fish target detection,counting and segmentation under the condition of fishery breeding,this paper collects video data in the breeding pond,converts,cleans and labels the data,and constructs a set of labeled data set.The data set includes daytime target detection data set,nighttime test set,daytime and nighttime sequence test set,which lays a data foundation for this research,and can also provide data sets for research in related fields.2)Aiming at the problems of underwater image blur and fish target resolution caused by turbid water in aquaculture environment,an enhanced hybrid fish target detection method is proposed.A variety of image enhancement methods are used to enhance the blurred underwater image.The enhanced image is respectively input into the fish detection model to obtain multiple outputs,and the multiple outputs are mixed.Then the non maximum suppression method is used to post process the mixed results to obtain the final detection results.The experimental results on YOLOv3,YOLOv4 tiny and YOLOv4 models show that compared with the detection results of the original image,the detection accuracy of this method is improved by 2.2%,8.3% and 1.4%respectively;The number of tests increased by 15.5%,49.8% and 12.7% respectively.At the same time,this method can effectively avoid the accuracy degradation of the target detection model on the underwater image enhanced by a single method,and achieve the purpose of improving the detection ability of the model.3)It is very important to monitor the nocturnal activities of fish in fish culture.Because the underwater image at night has different characteristics from the daytime,when the model trained by daytime image is used to detect the fish image at night,the accuracy is greatly reduced and the robustness of the model is poor.To solve this problem,a fish detection method based on day night field data migration is proposed.The labeled daytime training set is transformed into artificial night training set by image generation method,and then the daytime training set and artificial night training set are input into the model for training.The experimental results show that the accuracy of this method on the nighttime test set is improved by 10.26%,and the detection quantity is also greatly improved.It can improve the robustness of the model without investing additional data cleaning and labeling costs,so that the model is also suitable for nighttime fish detection.4)Due to the mutual occlusion and image truncation of fish in the shooting picture,taking the whole body of fish as the target for detection will produce a serious problem of missing detection.In order to improve the accuracy of fish quantity detection under aquaculture conditions,a method of fish quantity counting based on local optimization and improved output scale is proposed.By adding local information such as the head and tail of the detected fish,the largest number of the three categories is selected as the result of quantity statistics,so as to improve the accuracy of quantity counting.In addition,the whole body,head and tail of the fish are displayed as large-scale targets and mesoscale targets in the image.In order to improve the detection ability of the fish detection model for such targets,the output scale of the detection model is improved and the feature output of large-scale targets and mesoscale targets is increased.The experimental results show that the error between the detection quantity and manual counting on the daytime test set is small,and the accuracy is 96.3%.The number of detection on the daytime time series test set exceeds the number of manual counting,the accuracy rate is 88.9%,and the detection frame rate is 111 FPS.5)Segmenting fish targets from images is a key step in extracting fish biological information.Aiming at the low accuracy of fish segmentation in fuzzy underwater image,a fish image segmentation method based on target detection and edge support is proposed in this paper.Firstly,a complete contour extraction method based on target detection is designed.The fish target with complete contour is extracted from the image as the input of the segmentation stage,so that the whole image segmentation is transformed into the segmentation problem in the local region;Then,a Canny edge supported deep learning segmentation network is built to realize high-precision image segmentation of fish in the region.The experimental results show that the segmentation accuracy of this method is 81.75%,83.73% and 85.66% respectively on the models with VGG-16,Res Net-50 and Res Net-101 as the backbone network.Among them,the segmentation accuracy of the model with Res Net-101 as the backbone network is improved by 14.24%,11.36% and 9.45% respectively compared with the comparison models Mask R-CNN,U-Net and Deep Labv3.This method can be applied to fish object detection,fish quantity statistics and fish image segmentation in fishery culture.Based on this,a fish culture monitoring system can be developed,including real-time data browsing and auxiliary decision-making,so as to realize the automation and intelligence of fishery production.