Deep Metric Hashing Method For Large-scale Face Image Retrieval

Face image retrieval aims to identify images containing faces of the same person in the query face image.With the growing popularity of social networking on intelligent mobile services,the number of images and videos containing faces has witnessed an explosive increase.How to retrieve the target face quickly and accurately in the massive face database becomes an attractive research direction in the field of computer vision.The main challenges of face image retrieval are large intra-class variations and the great cost of time and memory.Therefore,it is significant to develop effective face image retrieval methods to address the above two problems.The retrieval performance of most existing hashing methods heavily depends on the feature they use.The traditional retrieval methods often use hand-craft features to represent the visual content of face image.However,these hand-craft features cannot well reveal the semantic information of face images,and often limits the performance of face image retrieval.In recent years,convolutional neural network(CNN)has shown its amazing performance in image classification,image recognition and other computer vision tasks.CNN features learned from images are more robust and can well capture the underlying semantic structure of images.The deep hashing method,which combines CNN and hashing algorithm,aims at learning face features with rich semantics and then mapping them into compact binary hash codes,which can improve the retrieval accuracy,reduce the storage space and shorten the retrieval time.Existing deep hashing methods usually have such problems as separating feature extraction and hashing encoding,ignoring the redundancy among hash codes,and only focusing on the pairwise similarity,which makes the training more difficult.To solve these problems,we propose a new Deep Metric Hashing(DMH)method for large-scale face image retrieval.This method incorporates deep learning,hash coding and metric learning into an end-to-end framework to learn discriminative and compact hash codes.The deep architecture and the supervised signals are collaboratively explored.Specifically:(1)In order to get a higher of hashing coding,we learn from Dense Net and build a network with dense convolutional block to extract multi-scale and robust features by connecting each layer to every other layer in a feed-forward fashion.(2)To reduce the redundancy among hash codes and make full use of the spatial structure information,a convolution and global-average-pooling module to generate compact hash codes and reduce a large number of the network parameters simultaneously.(3)Moreover,the Softmax Loss,the Center Loss and the Quantization Loss jointly guide the learning of the deep network by minimizing the prediction errors of the learned hash codes,which can lead to discriminative hash codes.(4)To evaluate the effectiveness of the proposed method,extensive experiments are conducted on two scalable face image datasets and experimental results exhibit the superior performance of our method compared with some state-of-the-art hashing methods.