Research On Image Retrieval Based On Deep Hashing Network

With the rapid development of the Internet and the popularity of mobile devices,the number of images online and offline has increased exponentially.Faced with the currently massive and complex image datas,how to find the image data one needed quickly from datasets is a major problem in the current image retrieval tasks.Recent years have witnessed increasingly attention on learning to hash methods for large scale image retrieval,and has achieved good retrieval accuracy in image retrieval tasks.However,due to the limitations of the worse capability for manual features to express images and the large errors caused by different hash functions when quantizing the feature vectors.There are still many improvements can be done in the current hash learning method.For most existing supervised hashing methods,the procedure of feature learning and hash code learning are separate,which may result in neither optimal hash codes nor discriminative feature representations.Considering the above shortcomings of the existing hash learning methods,this thesis proposes a deep supervised hashing network model for image retrieval tasks.The main research contents include the following aspects:1.Based on the pre-trained VGG16 model,the migrated network model is fine-tuned by using the image retrieve datasets.At the same time,we remove the last fully connected layer but add a hash learning layer and a weight learning layer to the network model,which constituted our end-to-end deep supervised hashing network to simultaneously perform feature learning and hash code learning.2.The pairwise labels are used as the supervised information in the training process of the network model,and the similarity between similar images is preserved to make the distance between hash codes learned from similar images through the network as small as possible.In the weight learning layer,we learn the adaptive weights on the corresponding bits of hash codes,so that the hash code of the shorter number can be directly filtered through the comparison of the weight sizes,thereby greatly saving the training time of hash codes with different number of bits.3.Propose an improved objective function to learn better hash codes in the hash learning procedure.By using regularization term and tanh-like activation function,the quantization errors caused by discrete optimization problem in network learning process are reduced,and the quality of hash code learned by network model is improved.