Image Representation Optimization Based On Deep Local Features

Instance-level image retrieval task has been attracting research interests in computer vision,and effective image feature extraction is the most important part.Deep convolutional neural networks(CNN)have recently shown their promise for generating universal representations in image classification and object detection tasks,this thesis will based on deep feature extraction,and propose several improvments for instance-level image retrieval task.Investigated previous deep feature extraction methods based on CNN,this thesis introduces a deep convolutional architecture based on local region detection,to enhance the robustness for matching of highly variable scenes in image retrieval task.This work yields invariant region-level descriptors by detecting specific regions in original image and handling the geometric transformations for images containing same objects or scenes,the final representation is able to leverage the strong feature expression of convolutional neural networks and address the lack of geometric invariance in previous approaches at the same time.Besides,in contrast to previous works employing handcraft local features or deep features extracted with convolutional neural networks(CNN),or utilizing pre-trained deep networks to generate image features,this thesis proposes a CNN fine-tuning approach based on ranking loss,which produces image representations in accordance with image retrieval task.Extensive experiments show that,on several standard retrieval benchmarks,the proposed approaches outperform previous approaches based on vector of locally aggregated descriptors or other convolutional neural networks based feature extraction methods.