| There are abundant resources in the ocean,and the exploitation of ocean resources is significant to people’s lives.Marine aquaculture is an important way of marine exploitation.With the gradual exploitation of marine aquaculture,low-efficiency manual way is gradually replaced by computer.With the application of computer vision and deep learning technology,refined mariculture based on vision has become possible.Through the use of machine-equipped cameras and visual analysis technology,the growth of cultured fishes can be monitored and the breeding process can be analyzed fine.This requires the computer not only to be able to identify various fish,but also to gather the parameter such as the status of the fish as body length etc.This article takes the underwater fish as the research object,based on the binocular vision theory,the target segmentation and detection algorithm are introduced to determine the specific position and contour size of the fish in the picture,and designs an automatic underwater fish body length.The automatic measurement method for the underwater fish body length estimation realizes the functions of classification,position detection and body length estimation of fish.The main research contents are as follows:(1)The basic theory and process of binocular vision technology are introduced to realize distance perception.The core of the binocular vision theory is the stereo matching algorithm,which calculates the disparity of binocular images pixels to determine the distance between the position represented by the pixel and the camera.This SGBM algorithm is adopted for stereo matching.(2)The object segmentation algorithm is introduced into the binocular visual distance calculation.Aiming at solving the problems of low accuracy and slow speed of the stereo matching algorithm for entire image,the image segmentation algorithm is used to obtain the contour of the fish in the binocular image for stereo matching algorithm.According to the constraint information of the fish,the local area is cropped according to the position of the fish contour in the image for matching calculation,which reduces the amount of matching calculation,reduces the calculation time,and improves the matching effect.At the same time,the stereo matching algorithm can only find the distance between the object and the camera.To calculate the length of fish body,the body length of the fish in the image needs to be estimated by the segmentation algorithm.The length of the fish in the picture is calculated by the contour of the fish,and the specific length of the fish is calculated by combining the length with the distance information obtained by the matching algorithm.(3)The object segmentation algorithm is mainly used for the identification of fish contours and cannot directly give the position information of multiple fishes.Based on this,a detection algorithm is introduced to detect and identify different types of fishes,and separate the location information of each fish.In order to reduce the amount of calculation.A multi-task joint network architecture of object segmentation and detection is introduced during the training process for fish detection,segmentation and body length estimation.In this paper,tensorflow is used to build a network model,and the object segmentation algorithm is integrated with binocular visual stereo matching calculation.Based on the MultiNet architecture,a network model based on multi-task,multi-object detection and segmentation is constructed to realize underwater fish detection,segmentation and body length estimation,and use the self-made data set and fish4knowledge data set to complete the model training.Experimental results show that the designed model and algorithm can complete underwater fish detection and body length estimation,and the estimation error of underwater fish body length is about 5%. |
