An Image Retrieval System and Method Based on Robust Deep Hash Network

Abstract

The invention discloses an image retrieval system and method based on a robust deep hash network, belonging to the fields of computer vision and pattern recognition. The invention introduces the space conversion network module into the deep hash network structure, so that the model can adaptively perform affine transformations such as scale scaling and rotation according to the picture content, and improve the generalization ability of the model. At the same time, the mutual learning strategy is adopted to increase the generalization and robustness of the network, and at the same time, it can also improve the training stability of the network and accelerate the convergence of the network. Using the relaxed hash generation function and improving the network structure to relax the binary hash code to a continuous real-valued space, the network can be derived and learned through the back propagation algorithm. This technical means can gradually approach the sign function after adding an amplification factor before the input according to the relaxed hash function, thereby replacing the sign function. Due to the removal of the regular term of the binary constraint, the introduction of redundant auxiliary variables and calculations is avoided.

Description

technical field [0001] The invention belongs to the field of computer vision and pattern recognition, and more specifically relates to an image retrieval system and method based on a robust deep hash network. Background technique [0002] In recent years, with the rapid development of Internet technology, massive pictures and video data are generated on the Internet every day. For example, the Tianyan monitoring system in my country's "Safe City Smart Community" plan generates tens of thousands of video files every day; the famous online picture and video sharing social software Sina Weibo has nearly 85 million videos and pictures uploaded every day , a total of more than 40 billion images were shared. Faced with such a large amount of image data, how to efficiently index the image data so that the target image queried by the user can be quickly and accurately retrieved from the massive image data has become a hot issue in the field of image retrieval research. Traditional ...

AI Technical Summary

Problems solved by technology

[0005] First, various loss functions have their advantages and disadvantages, resulting in insufficient network generalization ability and robustness

For example, a loss function suitable for classification tasks is not necessarily suitable for retrieval tasks; a comparison loss function suitable for retrieval tasks will bring about data imbalance and affect training efficiency; triplet loss functions need to be carefully selected during training. Appropriate triplets , limiting the improvement of network performance

[0006] Second, they are all faced with the non-differentiable problem of discrete space in deep hashing, that is, the image hash retrieval method of deep learning needs to use symbolic functions as activation functions to realize the operation of generating binary hash codes in the network. However, for any The gradient of the input sign function is zero, which makes the deep network based on the backpropagation algorithm impossible to train

Although the existing methods use binary constraint quantization loss to replace the symbolic function, the binary constraint quantization loss will change the distribution of features, which will affect the final retrieval results.

[0007] Third, the training set and test set of the deep convolutional neural network do not overlap, and there are certain differences in data distribution. The network that performs well on the training set often cannot obtain the same performance on the test set. Existing methods do not pay attention to how Make the network achieve satisfactory performance in both the training set and the test set, that is, the focus of the robustness of the network

However, the loss function of the existing deep hashing method is relatively single, and the advantages of each loss function cannot be combined. In addition, there is also the problem of non-differentiable discrete space, and the robustness and generalization of the network are poor.

Images