Research On Pattern Recognition And Application On Cattle Face Recognition Based On Deep Learning And Sparse Representation

Individual recognition based on image processing is a safe and non-aggression recognition method,which has attracted much attention from scholars both at home and abroad,and has made breakthrough progress.However,there are still two difficult problems in the application of image processing technology:(1)The data acquired from non-cooperative image individual recognition in the natural environment have many changes,the recognition algorithm needs to overcome the problems of occlusion,motion blur,out-of-focus,complex background,gestures and lighting changes and other issues.(2)Re-train the recognition model are necessary when there is a phenomenon of population fluctuation,such as population reproduction or purchasing new individuals.In view of the above problems,this paper takes Holstein cows as the research object,and proposes an incremental recognition algorithm framework.This framework makes full use advantages of convolutional neural networks features with good discriminability and strong mobility,and sparse representation classifier with fast computational speeds and easy addition features.Eventually,it achieves the purpose of real-time accurate incremental identification in a complex environment.This article mainly completed the following work:1.Based on deep learning features and sparse representation classifier,a high-performance real-time incremental recognition method was proposed.Firstly,extracted picture features using a well-trained CNN,then the sparse representation recognition model is constructed by utilizing above features,finally,the cattle faces are recognized according to the minimum residuals principle.When adding new samples or classes,just input them into the network,extract their image features and append them to the sparse representation dictionary for achieving incremental recognition.This method doesn't need to change the network structure or retrain the neural network.2.In order to meet the requirements of real-time identification,the idea of reducing the feature dimension as much as possible under the premise of ensuring algorithm's recognition accuracy is proposed to reduce the subsequent identification time-consuming.After fully studying the relevant knowledge of the convolutional neural network,theoverall training and single-layer training method were successfully used to automatically reduce the convolutional neural network features to 64 dimensions in the network training process.3.Compared the algorithm's accuracy and time consuming with CNN,CNN+SVM,SIFT,HOG+SVM and HOG+SRC in the non-incremental case.Experimental results show that the proposed algorithm can realize image recognition in complex scenes with the highest recognition accuracy rate of 99.94%.The time required for single head cow recognition is only 4.8 milliseconds.4.Research and verify the effectiveness of incremental recognition algorithm.Compared the recognition accuracy of algorithm when the newly added ratios are 500%,200%,100%,76.46%,50%,36.36%,and 20%.And,discussed the relationship between the number of registered pictures and the recognition accuracy rate under different newly added ratios,which can provide reference for practical applications.Taking the newly added proportion of 20% and the registered pictures number of 800 as an example,the accuracy of the individual identification of the algorithm can reach 99.70%.5.In order to apply this algorithm to the individual identification of dairy cows in the One Plus One Second Pasture in Wuzhong City,the Ningxia Hui Autonomous Region,we went to the pasture to carry out the collection of cattle face images.To objectively verify the efficiency of the image incremental recognition framework proposed in this paper and improve the automation degree of the algorithm,a set of automatic acquisition scheme for the image of the cow face is designed.And use this scheme to complete the cattle face video data collection in the farm environment.