Research And Implementation Of Animal Contour Segmentation Algorithm Based On Multi-scale Neural Network

Animal husbandry has gradually gone from extensive breeding to standardized management to "intelligent" farming.Using special measuring tools to measure animals is not only inefficient,but may cause animal stress.Therefore,people gradually use deep learning image processing to identify individual animals,eliminate animal stress reactions while acheieving contactless,long-distance and automation effects.In the case of light changes,similar color and environment on animal surfances,and blurred noise in images,etc.,they all brings great difficulties to target segmentation.In response to the above problems,this paper proposes an animal contour segmentation algorithm based on neural network,the main research contents are as follows:In this theiss,the fully convolutional network is selected as the basic network framework for animal contour segmentation,and the network is improved according to animal characteristics.In order to solve the problem that segmantic segmentation pixels of animal images are easy to lose under complex backgrounds,a multi-scal residual feature extraction module is proposed,and then replace 3×3 convolutions in the last three convolution blocks of the full convolution network,thus obatain more segmantic information in animal images and improve the ability to extract model features;Aiming at the problem of poor contextual informatiion correlation and misdivision of animal pixels when fully convolutional networks are connected,a RefineNet module is introuduced into the encoding part of the network,which uses RefineNet to fuse the characteristics of different dimensions,and finally generates highresolution segmantic feature diagrams,then improve the network segmentation accuracy.Finally,according to the particularity of animal data in animal husbandry,the Focal Loss functions is selected as the loss function in the process of networks training.Experiments on animal data sets show that the improved full convolution network can effectively reduce the impact of complex background on animal contour segmentation and improve segmentation accuracy.Aiming at the problem of blurring small target animal in the improved full convolution network,a segmentation model of integrated mixed attention mechanism is proposed.On the basis of the improved full convolution network,mixed attention is introduced to strengthen the extraction of important features and suppress information unrelevant to the target.At the same time,in order to solve the rough segmentation of the edges of animal contours,especially legs,ears and the other parts,CRF is used as global information to process the segmented animal image,so as to achieve the fine segmentation of the outline.Experimental results show that the proposed method can make the network model more accurate,and meet the purpose and requirements of the expected design of the model.