Method for searching images for vehicles on basis of convolutional neural network

Abstract

The invention relates to a method for searching images for vehicles on the basis of a convolutional neural network. The method comprises the steps that images sets of different vehicles are collected, training is conducted in a typical convolutional neural network model by taking the same vehicles as positive sample pairs and taking the different vehicles as negative sample pairs, similarity differences or classification errors are minimized, a group of vehicle feature expression methods are learned, a result of a data layer can be taken as texture features of the vehicles after the vehicle images are propagated forwards in the convolutional neural network model, the similarity between the features of the vehicles to be retrieved and the features of retrieve set vehicles is calculated through the features, and a result for searching the images for the vehicles is obtained by conducting sorting according to the similarity. According to the method, the vehicle image appearance expression methods are learned through the convolutional neural network, and compared with SIFT features, HoG features and the like, the purposiveness is higher, the features are more visual, the extra metric learning process is not needed, the searching accuracy and precision are significantly improved, the feature dimension number can be controlled to be at the small magnitude, and quick searching in a large image library can be achieved.

Description

technical field [0001] The invention belongs to the technical field of intelligent transportation, and relates to a method for searching cars by image based on a convolutional neural network. Background technique [0002] The image search method relies on vehicle appearance features to sort the search set by similarity, and can find all images containing the vehicle that the user is looking for from the video or image set. The search process does not depend on the license plate information. The same applies. [0003] Convolutional neural network is a hot spot in current research and industrial applications. Compared with traditional artificial intelligence algorithms, such as neural networks and support vector machines, deep learning algorithms can greatly improve the accuracy of image classification, and have already been used in the field of face recognition applications. Accuracy beyond the naked eye recognition. [0004] There are many ways to express vehicle appearanc...

AI Technical Summary

Problems solved by technology

[0005] In these existing vehicle appearance representation methods, manual methods such as SIFT and HoG are commonly used to calculate texture features. This traditional manual feature extraction mode is suitable for most pattern recognition applications and has good versatility. However, due to the fixed calculation method, it cannot achieve Very high accuracy, which is not conducive to the development of vehicle retrieval

Images