Morphological Parameters Measurement Based On Deep Learning For Marine Fisheries Breeding

Marine resources are very abundant in China.Marine aquatic products are rich in nutrition and have high commercial value.The aquaculture industry is developing rapidly.Aquaculture increasingly requires better fanning strategies to improve production quality and efficiency.Large scale breeding and high quality breeding need a fast and high flux measurement technology to provide data support.Related fields of marine scientific research also need to measure some morphological parameters of fish and shrimp.Therefore,this paper selects prawns as the main object of the research and uses computer vision technology to measure the morphological parameters of prawns.Focusing on the goal of measuring from the back image and the profile image of the prawn,the specific research content is as follows:First,after twice data collections,the prawn front image data set and the prawn side image data set were constructed respectively.A prawn back image dataset was constructed,which contained 219 Litopenaeus vannamei,190 Metapenaeus ensis,243 Marsupenaeus japonicus,totaling 1956 images.A prawn profile image dataset was constructed,which contained 290 Litopenaeus vannamei,with a total of 1160 images.Secondly,for the target of measuring carapace length,carapace width,and body length from the back image of prawns,this paper proposes a measurement method based on Mask R-CNN and a measurement method based on CPN from the perspectives of image segmentation and keypoints estimation.Experiments show that these two methods have relatively small measurement errors and do not require preprocessing steps for prawn posture correction.Then,this paper studies the main skeleton extraction method and builds the Sk-Unet model to extract the main skeleton of the prawn.Sk-Unet uses the basic structure of Unet,uses dilated convolution and multi-branch structure to enhance the feature extraction,and uses the CBAM module for Attention operation,and finally combines the features of all block layers.Experiments show that Sk-Unet can obtain better results with relatively few parameters.Finally,due to the loop problem of the main skeleton,it is impossible to extract the main skeleton lines correctly.In this paper,we propose the solution of the end-point open loop and extraction of main skeleton lines.Select main skeleton’s length,main skeleton’s Fourier descriptor,area,perimeter and area perimeter ratio as features,and use regression prediction model to predict body length.Experiments show that the linear regression model works best.