Method Research On Fine-grained Image Classification Of Birds Local Features

In recent years,with the rapid development and maturation of the intelligent image classification technology,the effects and performance of general image classification algorithms have gradually surpassed the human visual system,and related research results have inspired people to challenge higher-level issues,then the fine-grained image classification problem began to attract the attention of researchers,and this classification task needs to distinguish multiple sub-categories of the same major category.However,the samples between different sub-categories have high similarity,and the samples under the same subclass will have great differences due to interferences such as shooting distance,target pose,and complex background.Thus,the task of the fine-grained image classification is far more difficult than the ordinary image classification.The key to solving the above problems lies in the localization and extraction of local features,because for the fine-grained images,the slight differences between the samples and the commonness of the intra-class samples that are disturbed are hidden in the local area of the image,which can only be described accurately by stable and detailed local features.Therefore,the key problem of the fine-grained image classification is how to extract more effective and robust local features.To solve the key problem of the fine-grained image classification tasks,this thesis analyzes the advantages and disadvantages of the two types of feature extraction methods,including Convolutional Neural Network(CNN)and Scale Invariant Feature Transform(SIFT).And the idea of local features combination is proposed in order to improve the accuracy of the fine-grained image classification.The main work include:(1)An idea of local feature combination based on CNN and SIFT is proposed.Through combining the SIFT feature which has the description of local key points of the image and the CNN feature which has the acquisition of high-level image semantics,the strong and distinctive description for fine-grained images is obtained.The more powerful feature representations are used to improve the accuracy of the fine-grained image classification.(2)According to the sample characteristics in the fine-grained classification task,the general SIFT feature extraction process is optimized to enhance its orientation invariance.Then,SIFT features before and after optimization are used for classification based on feature coding method,and multiple sets of contrast experiments are performed under different feature coding parameters.The effect of the selection of feature coding parameters on classification is illustrated by analyzing the experimental results,and the optimization effect is proved at the same time.(3)Considering the importance of local features for the fine-grained image classification tasks,the CNN feature extraction scheme is designed,and the final scheme is analyzed and determined based on the experimental results.(4)The realization of multiple local features combination based on SIFT and CNN features is performed in this thesis,and experiments on the fine-grained classification standard data set CUB-200-2011 demonstrate the effectiveness of the method.