| Image retrieval based on feature representation and matching plays an important role in computer vision.The current image feature representation has shown good results in coarse-grained image retrieval through hand crafted or deep learned features,but it is still difficult to satisfy actual needs in fine-grained image retrieval.The main challenges and difficulties lie in that the visible differences between different sub-categories of images under the same meta-category are very subtle,while there may be large visual differences between the same sub-category images.In this case of large intra-class variance and small inter-class variance,it is difficult to accurately capture the fine-grained differences between images with manually extracted features.For deep learning,one has to extra collect a large amount of labeled data.To avoid the high cost of annotation,it has become urgent to break through the performance bottleneck of hand-crafted feature extraction in fine-grained image retrieval.Towards this end,this thesis studies fine-grained image feature extraction and representation based on pre-trained convolutional networks from the perspective of feature transfer,and conducts image retrieval experiments for justification.The main works of this thesis are as follows:(1)Based on pre-trained convolutional networks,the current fine-grained feature extraction methods mainly select salient regions from convolutional feature maps,while the differences between convolutional channels are not considered.This thesis proposes a fine-grained feature extraction module,which can select those salient regions from convolutional feature maps and meanwhile remove the non-salient convolution channels,thereby enhancing the discriminativeness of selected fine-grained features.(2)Fine-grained features with salient selection can localize object parts in the image,but ignore the global context information.This thesis leverages a coarse-grained module to obtain the global image feature representation.The coarse-grained and finegrained image features are then fused to form an improved fine-grained image representation through a weighted concatenation.(3)In order to justify the advantages of the proposed fine-grained image representation,this thesis conducts extensive experiments on six popular benchmarks for fine-grained image retrieval.The results demonstrate that the proposed method achieves significant improvement over existing state-of-the-art methods.Besides that,this thesis also discusses the setting of hyper-parameters in the proposed method,and verifies the rationality of the design of the proposed method using ablation experiments. |
