Traffic image retrieval method based on depth learning

Abstract

The invention provides a traffic image retrieval method based on depth learning in an intelligent traffic application scene for performing traffic monitoring video image retrieval through depth hash encoding. The method comprises the following steps: dividing a target data set into a training set and a test set; acquiring features and image hash codes of a target category and a target color by a depth convolutional neural network; optimizing a hash function by back-propagating of classification loss and hash encoding loss of the features of the category and the color; performing hash encoding on an image by the hash function, and calculating and inquiring a Hamming distance between the image and the has code of the image in the test data set to represent the similarity of the images; and performing similarity score sorting according to the size of the Hamming distance to retrieve images. According to the method provided by the invention, rich multistage semantic information in the images is retained in the image retrieval, the special attribute information of targets in the images is utilized, multiple tasks of retrieval and image attribute classification are accomplished by sharing a network structure, and the retrieval is assisted by a classification task.

Description

technical field [0001] The present invention relates to the technical field of content-based image retrieval, in particular to a deep learning-based multi-attribute deep hash coding image retrieval method, specifically a method for realizing traffic image retrieval based on deep learning. Background technique [0002] With the advancement of projects such as safe cities and intelligent transportation, surveillance cameras have left image data for most cases, which has brought great convenience to the police in solving cases. However, the work of finding and analyzing videos often consumes a lot of time and manpower. The need to find relevant information more easily in massive videos is getting stronger and stronger, and video retrieval technology is becoming more and more important. The amount of traffic surveillance video data is very large, and it is not easy to find specific targets in such massive data, so image retrieval in this traffic application scenario came into b...

AI Technical Summary

Problems solved by technology

Using the image hash method can quickly calculate the Hamming distance between hash codes and greatly reduce the amount of data storage, but its disadvantage is that it depends on the extraction of image features, and the complexity of an image often cannot be determined by several features. expression, at this time the encoding function cannot be well close to the original image information, and the retrieval accuracy will decrease

Although this method handles the high-level semantic information contained in the image, it does not use the attribute feature information of the object in the above image. This single deep hash coding method still has room for improvement.

Images