Design And Implementation Of Deep Piecewise Hashing For Image Retrieval

With the growing of image data on the Internet,the demand for image search is increasing.While image hashing algorithm is widely used due to its advantages such as fast computing speed and less storage space,so many excellent algorithms have emerged.However,there are still some problems to be solved in these algorithms.First of all,the traditional data-based hash method has strong universality and fast operation speed,but it is not advantageous in the accuracy of results.Then,although the image hash algorithm based on convolutional neural network which is widely used now has a high accuracy,it needs to train a special hash generation network,which requires a lot of training time and storage space.Moreover,many image hashing algorithms using neural networks are often unable to determine what information is encoded into the hash code,which has great semantic uncontrollability.For the above problems,the deep piecewise hash algorithm gives the corresponding solution.First,to solve the problem of low accuracy of traditional hash method,the algorithm gives the task of image feature extraction to convolutional neural network,but traditional image processing method is used for feature processing,which preserves both accuracy and computing speed.Then,for the problem of large training time and space requirements of convolutional neural network,the network required by the algorithm directly adopts the classification neural network,which does not need to train the network specifically for the hash algorithm,and realizes the sharing of network models,thus reducing the network training time and storage space from the perspective of sharing.Finally,regarding the problem of the uncontrollable semantics of hash coding,the algorithm adopts the piecewise strategy to segment the whole hash code according to different semantics,and uses the classification neural network and the significance algorithm to encode the category,feature and spatial information of the image into the hash code,so as to realize the semantic management of hash code.After algorithm implementation in the three public data sets of MNIST,CIFAT-10 and AwA,it is proved that the algorithm proposed in this paper achieves good results in precision and mean average precision(MAP).