Research And Implementation Of Fine-grained Bird Recognition Based On Deep Learning

Along with the development of deep learning and the improvement of computer hardware performance,the general image recognition task in the field of computer vision has made great progress,whereas fine-grained visual classification is aimed at classifying the subordinate-level categories under a basic-level category,due to its characteristics of high intra-class variances and low inter-class variances,it is hard to obtain accurate classification results only by the state-of-the art Convolutional Neural Networks so that fine-grained visual classification has become a hot research topic.Birds have an intimate relationship with the natural ecology and human living,accurate recognition of birds is of great significance to related practitioners and bird lovers,therefore,identifying birds effectively from the perspective of computer vision is a valuable and promising but challenging task.The manually constructed birds image dataset and the common fine-grained visual classification dataset CUB-200-2011 are selected as the research objects.After studying and using some common data augmentation methods to expand the training dataset,necessary data preprocessing is carried out on training data.Then,three different image recognition algorithms are used to conduct experiments on the datasets.Firstly,the method of fine-tuning on pre-trained model is tried,and the results are used as the baseline of the task in this thesis.Secondly,a recognition method based on saliency maps is proposed,which can get the cropped images of key parts,input them and the original images into two convolutional models respectively to make decision on final results jointly.Finally,we propose a data augmentation method based on the class activation mapping,which can generate specific images for training,including cropping,dropping and zoom sampling of the attention region.The algorithm can effectively enhance the feature learning ability of the model.After training a model with high accuracy and through demand analysis and overall design,a browser-based fine-grained bird image recognition system is developed and implemented,which provides the functions of uploading,recogniting and analysising,data storaging,log collecting and so on,and can make a great user interface.This thesis explores a deep learning algorithm of fine-grained birds image recognition which has higher classification accuracy only with the need of image-level annotations.After repeated experiment and analysis,the CAM-guided data augmentation method reaches the accuracy of 88.2%which is closer to the existing advanced algorithm,and it has a better efficiency and interpretability.This kind of data augmentation method of generating more effective images also opens up a new research direction for fine-grained birds classification.