Research On Freshwater Fish Species Recognition And Length-mass Estimation Based On Underwater Stereo Vision

In aquaculture,obtaining information on the species and growth of cultured fish is important for accurate feeding,harvest grading and culture management.The mass of fish is the most intuitive information to reflect the growth information.Traditionally,the mass of fish is usually obtained by manual weighing after catching,which is not only time-consuming and laborious,but also causes stress to the fish and affects the normal growth of fish.How to obtain the mass of fish efficiently and conveniently is of great importance to fishery farming.With the rapid development in the field of machine vision,this paper combines binocular stereo vision and deep learning to obtain the mass of the fish by using the following methods:Micropterus salmoides,Carassius auratus,Ctenopharyngodon idella,Channa argus,and Silurus spp,five common freshwater fish were studied for the recognition methods of species and length-weight estimation of free-swimming fish underwater.The research results are as follows.（1）The Key Ponit R-CNN network,which can accomplish species recognition and key point detection simultaneously,was constructed based on the Faster R-CNN network by adding the key point detection branch.Underwater images of five species of freshwater fish,namely Micropterus salmoides,Carassius auratus,Ctenopharyngodon idella,Channa argus,and Silurus spp,were collected underwater by an underwater image acquisition platform,and the images were labeled to complete the construction of an underwater image dataset of freshwater fish.m AP of object detection at Io U=0.5～0.95 after the Key Ponit R-CNN network was trained using the data-enhanced underwater image dataset of freshwater fish was 89.4%,and the accuracy of species recognition can reach 100%by screening out the detection results with low species confidence;the m AP of keypoint detection is 94.6%at Io U=0.5～0.95.（2）A binocular vision measurement algorithm was written by reconstructing the points in the image in three dimensions by binocular vision and calculating the length between two points using the Euclidean distance formula.The effect of underwater image distortion on the length estimation was analyzed and the length results were corrected accordingly.The results show that the average relative error of the measurement algorithm after image distortion correction is 1.92%,which is 1.58%lower than the measurement error of the original image.The Key Ponit R-CNN network detection was combined with binocular vision measurement algorithm to write an underwater freshwater fish body length estimation algorithm.The underwater freshwater fish length estimation algorithm was validated through experiments.Compared with the artificial measurement results,the average relative errors of body length estimation of Micropterus salmoides,Carassius auratus,Ctenopharyngodon idella,Channa argus,and Silurus spp were 5.6%,7.1%,6.0%,5.8%and 5.8%,respectively.（3）The body length and mass information of Micropterus salmoides,Carassius auratus,Ctenopharyngodon idella,Channa argus,and Silurus spp were collected,and the mass estimation models of these five species were constructed by using the power function regression model(2=(6⁽（7）,where(2 is the mass of the fish（g）and L is the body length of the fish（mm）.The mass estimation models(2=1.751×10^-63.502（²=0.9877）for Micropterus salmoides;(2=1.737×10^-53.141（²=0.981）for Carassius auratus;(2=3.755×10^-52.89（²=0.9956）for Ctenopharyngodon idella;(2=1.036×10^-53.056 for Channa argus（²=0.9991）;and(2=1.153×10^-63.372（²=0.9963）for Silurus spp were established.Then,these model of five modles were combined with species recognition and underwater freshwater fish body length estimation algorithms to build underwater freshwater fish mass estimation algorithm.The underwater freshwater fish mass estimation algorithm was also validated through experiments.Compared with the artificial weighing results,the average relative errors of mass estimation of Micropterus salmoides,Carassius auratus,Ctenopharyngodon idella,Channa argus,and Silurus spp were 17.8%,15.4%,14.6%,16.2%and 17.5%,respectively.