Research On Hash-based Image Retrieval

With the advent of the Internet era,the promotion of large data technology,the development of multimedia technology and the generation of massive image and video data,people are increasingly demanding for image retrieval.Now the image retrieval is characterized by an increasing amount of data,the data dimension is also getting higher and higher.Traditional tree-based data search algorithms degrade in the current retrieval task.In the high-dimensional data search tasks,the algorithm based on the hash tends to obtain more excellent search results.This paper studies the hash algorithm in image retrieval.The quantization and coding method in the traditional hash algorithm is improved by introducing Euclidean distance into the measurement methods of hash code.A depth hash model combining depth learning and hashing is proposed.The main work of this paper is as follows.A double-bit embedded hash algorithm based on Euclidean distance is proposed.In the quantization stage of the traditional hash algorithm,only one bit of hash code is generated on each projection,which will cause large error.While the Euclidean distance-based double-bit embedded hash algorithm in the quantization phase,each projection will produce two bits of the hash code.Euclidean distance is used as the metric of hash coding,which improves the ability of hash coding to maintain the original spatial relation.Aiming at the slow computation of Euclidean distance,a Euclidean distance calculation method based on bit operation is designed to improve the computing speed of Euclidean distance of double-bit embedded binary code.At the same time,the Manhattan hash algorithm based on bit manipulation is proposed to improve the search speed of double-bit embedded Manhattan hash under the premise of keeping the same precision.A depth hash model based on class distance constraint is proposed,which is an end-to-end convolutional neural network model combined with depth learning algorithm and hashing algorithm.The superior performance of convolution neural network in the high-level semantic representation of pictures makes it possible to achieve better results in image retrieval for semantic information retrieval.Deep Hash Model based on category constraints adds category information layer to measure the semantic similarity between different categories on the basis of popular depth hash model,and the loss function can be used to increase the distance between easily confusing categories.Three new constrained functions are also proposed in the model,class distance constraint function,minimum quantization error constraint function and class distance adaptive constraint function.They can effectively guarantee the optimization of network performance.