Semantic Part Segmentation Method Of Chiloscyllium Plagiosum Based On Encoder-decoder Network And Its Application

Intelligent monitoring and video behavior analysis have become important technical means for observing,recording and quantifying behavior patterns in animal behavior research.The chiloscyllium plagiosum has a high economic value and medical research value.However,under a real artificial breeding conditions,marine fish have high requirements for the breeding environment such as environment and temperature and other factors,which often cause large-scale disease and death.The use of video images to quantify and analyse fish movement behaviour can help identify abnormalities behavior and provide early warning,which will effectively improve the level of conservation of farmed objects in aquaculture.In this paper,aiming at artificially domesticated chiloscyllium plagiosum,a semantic part segmentation method based on deep neural networks is proposed,and the segmentation results obtained from this method are applied to profiling the body movement posture of chiloscyllium plagiosum,and then identify the fish body movement in the video sequence.The main research contents of this article are as follows:(1)The semantic segmentation method based on Encoder-Decoder deep network is studied.Early semantic segmentation methods rely on hand-designed features for pixel classification,which is time-consuming and labor-intensive and has poor robustness in dealing with interference from complex environments.In this paper,a convolutional neural network method based on deep learning is used for semantic segmentation.Based on the detailed research and analysis of the principles of deep network architectures such as FCN-8s,SegNet and DeepLabv3+,we try to use these deep networks,which are mostly used to segment different target objects in the scene,to semantically divide different body parts of the same target object in the scene.(2)Research the deep small-sample learning method for semantic segmentation of fish body parts,and a data sample set of chiloscyllium plagiosum body images were established.Firstly,the chiloscyllium plagiosum is divided into seven visible body part components based on its morphological features;then the 476 chiloscyllium plagiosum sub-images extracted from the panoramic farming surveillance video are processed to carry out pixellevel labelling of each part,with these data eventually reaching a total dataset of 1,944 through data enhancement to establish a semantic dataset of fish body parts,with 1,166 images in the training set and 778 images in the test set.This work has laid a solid foundation for further in-depth analysis of shark video behavior.(3)A comparative experimental study on the segmentation of the semantic parts of the striped bamboo shark body was carried out.The pre-processed training set was fed into semantic segmentation network model to develop an optimum FCN-8s,SegNet and DeepLabv3+ model by the fine-tuning of the network parameters using the deep learning framework.In this paper,We conducted comparative experiments on fish part segmentation based on three deep network models,respectively The results on the test set showed that the segmentation method achieved a complete segmentation of the part of chiloscyllium plagiosum,the algorithm based on DeepLabv3+ network is the best,the accuracy of the head,right pectoral fin,left pectoral,right ventral,left ventral,trunk and tail areas are reached by 0.97,0.91,0.92,0.91,0.91,0.90,0.98 percentage point,respectively.(4)Futher study the application of how to use the segmentation results of the semantic parts of the fish body.In the video sequence images,each body part of the target object must satisfy the structured layout spatially,and this layout will show a statistical change regulation in the time sequence images.Because it can effectively distinguish and locate the visual parts of the fish body,by locating the torso and the center of mass of the fish head,the body-based coordinate system of the fish body is established,and then according to the direction change of the body-based coordinate in the video sequence image,it can be determined that the fish body is turning left or right.The experimental results show that the semantic part segmentation results can effectively discriminate the action posture of the chiloscyllium plagiosum body targets,which can lay the foundation for the recognition of abnormal fish behavior and further development of animal behavior experiments for the chiloscyllium plagiosum.