Deep Learning Based Feature Representation And Object Retrieval

Content-based image retrieval technology has always been a popular research direction. Nowadays, the CBIR technology can help people get the information they want more quickly and accurately, under the background of explosive growth of Internet pictures and video resources.Deep learning is very hot in recent years. Due to the Strong feature learning ability and nonlinear feature representation ability, deep learning has achieved a huge improvement in computer vision. In this paper, we mainly study the feature representation based on deep learning and its application in image retrieval.An effective image retrieval system is divided into two parts: one is a powerful image representation; the other is a fast retrieval process. This paper focuses on these two aspects. In the aspect of feature extraction, this paper firstly studies the difference of expression ability between different layers in CNN, and then proposes a multi-level pooling method and multi-layer feature fusion method, and finally filters the feature by using PCA or LDA dimensionality reduction method. In the aspect of fast retrieval, this paper proposes a hash algorithm based on deep learning, which combines feature learning and hash function learning with convolution neural network.In particular, the main work of this paper is summarized as follows:The difference of feature representation in different layers of convolutional neural network is studied, and it is found that convolutional channel features are more effective for image retrieval compared with the feature of fully connected layer.A multi-level pooling method is proposed for feature extraction,which combines the local features in the convolution feature maps with the global features of the whole map to make the feature more robust;A method of multi-level feature fusion is proposed, which improves the generalization ability of the final feature by combining the low-level visual features and the high-level semantic features. Experiments show that the multi-layer fusion method improves performance in multiple datasets.A region-aware deep hashing algorithm is proposed. It's composed of region proposals extraction algorithm, RoI pooling,multi-label classification loss function and weighted triple loss function, which enable the network to simultaneously learn feature representation and hash function.