A car recognition method based on pooled multi-scale deep convolution features

Abstract

The invention discloses a car model recognition method based on pooled multi-scale deep convolution features. Firstly, for each car model image in the car model database, its deep convolution features are extracted according to different scales, and the first scale is not processed; The depth convolution features of each scale are subjected to PCA dimensionality reduction; then the local feature aggregation descriptor encoding is performed; then the dimensionality reduction is performed again through PCA to obtain the feature representation of the current scale; the features of all scales are cascaded and pooled to obtain the final image of the current image The feature representation of the car model image is used for linear support vector machine training to obtain a car model recognition system; the vehicle to be recognized is also obtained from its feature representation, and then imported into the recognition system to identify its car model. Traditional deep convolution features lack geometric invariance, which limits the classification and recognition of vehicle models in variable scenes. The present invention adopts image pooling multi-scale depth convolution features to solve this problem well, and has a high practicality and robustness.

Description

technical field [0001] The invention belongs to the technical fields of image processing, pattern classification and recognition, and in particular relates to a vehicle recognition method based on pooled multi-scale deep convolution features. Background technique [0002] Traditional vehicle recognition technology includes vehicle detection and segmentation, feature extraction and selection, pattern recognition and other processing. This type of technology faces many difficulties: how to segment the complete target vehicle area in a complex background is the premise and basis of vehicle type recognition; how to select representative features from among the many features of the car and convert them into effective parameters It is also extremely important; after obtaining the characteristic parameters, how to correctly select and design the classifier also directly affects the accuracy of the final recognition. [0003] The concept of deep learning originated from artificial ...

AI Technical Summary

Problems solved by technology

However, it does not perform well in car model recognition, and the global convolutional neural network features lack geometric invariance, which limits the classification and matching of variable scenes.

Images