Research On Key Techniques Of Sika Deer Individual Recognition Based On Computer Vision

As a special economic animal,the sika deer has excellent medical,dietary,and health care effects in various parts,which has been welcomed by consumers at home and abroad for a long time.With the gradual expansion of breeding scale and the development of information technology,the sika deer industry in China is gradually developing from free-range breeding to informatization,intelligence and precision.Accurate and efficient identification of sika deer individuals is a prerequisite for intelligent breeding in deer farms,which is helpful to track and analyze the physiological,nutritional and health conditions of sika deer and lay a foundation for further product traceability.Traditional individual identification methods for sika deer mainly include artificial observation,ear carving,and radio frequency identification techniques.However,traditional identification methods have the disadvantages of low efficiency,large errors,and harm to sika deer.In recent years,the development of computer vision technology has made image based non-contact recognition replace traditional image recognition technology as a mainstream method of animal individual recognition,and has been widely used in the field of animal recognition.But the traditional machine vision needs to extract features manually,and the recognition model has a large number of parameters,which is difficult to deploy to the real farm with limited computing power.To address the above problems and the current status of research,this thesis carries out research work on the identification of individual mergansers by analyzing images of mergansers with complex backgrounds collected in real farm environments and using them as research samples.According to the characteristics of the sika deer dataset,two convolutional neural network models were improved and optimized.After that,the advantages of the two models were extracted and enhanced to design the third model.The three models and other classical models are trained and validated with the sika deer spot dataset and the sika deer face dataset,as well as the performance comparison.The main research work of this thesis is as follows:(1)Build sika deer data.Collect the spotted images and facial images of sika deer in a real deer farm farming environment.Using object detection to locate the coordinates of deer face in the sika deer face image,and using the method based on Open CV to extract the deer face according to the coordinates to form the sika deer face dataset.Perform data enhancement,preprocessing,and other operations on each sample data to improve the applicability of the model.(2)Sika Deer Individual Recognition Based on Multi-Light Algorithm.A lightweight Multi-Light recognition model was constructed to address the problem of large parameter quantities and low recognition accuracy in CNN models,in order to achieve individual identification of sika deer in complex backgrounds.The model uses multi-scale convolution to extract multi-scale feature points,and utilizes the characteristics of residual structure to ensure that the extracted features are not lost.Finally,SE attention mechanism is introduced into the fused output structure to increase the nonlinear ability of the model.The recognition accuracy of the Multi-Light model on the deer spot dataset is 96.68%,and the recognition accuracy on the deer face dataset is 93.51%,the model size is 3.61MB,and the model parameter amount is 9.4×10~5.(3)Improved high-precision G-Res Net model for identifying individual sika deer.To solve the problem that the facial features of sika deer are not obvious,which leads to low recognition accuracy of CNN model,a high-precision G-Res Net recognition model based on Ghost and dilated convolution optimization is constructed.The cheap linear structure in Ghost is used to connect the feature flow between bottlenecks,reduce the number of parameters with dilated convolution,and focus on more important features with attention mechanism.Finally,the G-Res Net model has an accuracy of 96.25%for recognition of individual deer on the deer face dataset,which is 2.74 percentage points higher than the lightweight Multi-Light model,and the size of the G-Res Net model is 6.51 MB.In comparison with other classical CNN models,the G-Res Net model has high recognition accuracy while taking into account the low number of parameters.(4)Efficient identification of individual sika deer by constructing Two-Branch model with feature fusion.To further improve the recognition accuracy of individual recognition of sika deer,the light-weight Multi-Light model and the high-precision G-Res Net model are feature fused to build a Two-Branch model with double branch feature fusion,and the two branch models are further optimized separately to highlight the respective advantages of the two branch models.The G-Res Net model of Branch2 is used to achieve high-precision recognition,the Multi-Light model of Branch1 is used for feature calibration,two different branch models in the Two-Branch model are used for feature extraction of the sika deer,and finally the feature fusion output is realized to improve the recognition accuracy of the individual identity of the sika deer.Finally,the Two-Branch model is trained and validated on the plum deer spot and sika deer face,and the recognition accuracy is 98.86%on the plum deer spot dataset and 98.42%on the deer face dataset.The results of the above research and experiments show that the sika deer individual recognition method based on computer vision proposed in this paper can accurately and quickly identify the sika deer individual identity from multiple perspectives,and has certain practical value,which is helpful for further practical application of sika deer individual recognition and provides a theoretical basis for other animal individual recognition.