Super-resolution Reconstruction And Recognition Of Low Pixel Cow Head Images

The individual animal identification has played an important role in production management,control of disease outbreaks,vaccination and product traceability.And image identification is one of the core technologies of intelligent livestock breeding.In cow farm monitoring applications,cow identification accuracy is seriously affected by the higher position of camera,long distance between camera and cow and low camera resolution.This paper attempts to recover cow head identification information by combining a superresolution network with an identification network and improve cow identification performance.The specific works are as bellows.(1)The head images of Holstein cow in farms were collected from various angles and various environments using a variety of devices,including mobile phones and digital cameras.The cow head identification data set was constructed after screening,cow head extraction,blurring image cleaning,similar image cleaning,and ear tag blurring,etc.The image processing efficiency is greatly improved by the developing of semi-automated cow head extraction and ear tag blurring tools,and make it possible to process large amounts of cow head image data.The dataset contains 85,200 annotated cow head images from 1000 Holstein cow individuals,which is enough for deep learning training.(2)Tested the super-resolution reconstruction performance of three classical superresolution models at different magnifications on low-resolution cow head images,and constructed a high-performance super-resolution model that achieves better performance of PSNR than RCAN network by introducing an adaptive multiscale up-sampling module on the recovery of cow head images.In order to recover larger images with limited memory space,a lightweight super-resolution model is proposed with a feature extraction network of only 24 layers of convolution in the backbone and only 64 channels in the feature vector,increasing the number of convolution kernel before the activation function,while replacing the activation function with PRe LU,introducing adaptive residual connection and adaptive multiscale up-sampling module to maximize performance with a limited number of parameters.(3)A number of current mainstream recognition networks were trained and tested using the cow head dataset constructed in this project.The degradation of recognition performance of recognition models at different training resolutions and test resolutions was also analyzed.(4)A low-resolution cow head recognition framework is constructed that combines the super-resolution model with the recognition model to reconstruct the cow head image recognition information,and an alternate training approach is introduced to ensure smooth convergence of the network framework.An identification accuracy of 95.46% was achieved on small size(14 × 12)cow head,which is a significant improvement in identification performance compared to the 46.94% accuracy of the Bicubic interpolation method.