Underwater Visual Intelligence Based On-Spot Fish Size Measuring System

The aquaculture industry has made important contributions to China’s economy,employment and exports.At present,the vast majority of aquaculture units adopt the pond culture method,and use the traditional methods of manual fishing and measurement to monitor the growth of cultured fish.This manual method is not only difficult to obtain sufficient data,but also consumes considerable human resources,and is prone to damage the growth of cultured fish stocks.Driven by the above problems existing,it has become the key demand of aquaculture units to obtain the growth information of cultured fish efficiently,massively and accurately in a non-contact way leveraging underwater visual intelligence.This thesis focuses on the non-contact measurement of fish size based on underwater visual intelligence.The main contents are as follows:(1)As the preprocessing of input images,the blur problem of underwater image in the imaging process and the calibration of stereo imaging system are solved.An underwater image clarify algorithm is implemented.Through the image clarify model based on Swin IR structure,the high-definition restoration of underwater image is implemented.Also the calibration of binocular camera is realized,and the binocular image is calibrated by the parameters to make it comply with epipolar constraints.The clarified and calibrated image is used as the input of the follow-up visual intelligence based size measuring module,which is helpful for the subsequent modules to extract more accurate information and obtain more reliable monitoring results.(2)Aiming at the problem of non-contact measurement of underwater fish size based on visual intelligence,a fish size measurement method combining binocular vision,object detection and key point detection is proposed to realize low-cost and efficient non-contact fish size measurement.In the proposed method,based on sufficient data acquisition and training,through the faster R-CNN object detection the non-contact fish positioning problem is solved.And through the Stacked Hourglass network with position constraints the key-point detection is realized.According to the imaging principle of binocular camera,the key-points of the fish body are re-projected into 3-D space,and the fish size is calculated according to the key-point coordinates in the 3-D space.Experiments show that the proposed method has high size measurement accuracy,good adaptability in different scenarios,and is easy to deploy.It can be used as the optimal scheme for aquaculture farms to optimize human resources and improve aquaculture efficiency.(3)Aiming at the problem of underwater motion monitoring of fish,an object tracking algorithm leveraging image deep feature is proposed to realize the object tracking problem upon fish shoal,and further improving the accuracy of fish size measurement according to the fish object tracking results in continuous frames.In the proposed algorithm,the object in continuous frames is located through the detection results of object detection network.The motion trajectories of objects are predicted with SORT algorithm,and the consistency of the tracking object is distinguished leveraging the intermediate features of the key-point detection network.After the object tracking results of continuous frames are obtained,the median in the nearest period is taken as the size measurement value of the same object,so as to obtain more reliable results.Experiments show that the proposed object tracking method can accurately track multiple objects in continuous frames,and the fish size measurement error after object tracking correction can be reduced.