Research And Implementation Of Fine-Grained Classification Algorithm For Vehicle Color Based On Deep Learning

The vehicle color recognition technology is of great significance for vehicle identification and the development of the intelligent transportation system.However,some problems in vehicle color recognition technology,such as the small size of dataset,small variety of colors,and the effects of illumination and extreme weather in the environment,have caused the low recognition accuracy of traditional methods and the inflexibility of practical use.In order to solve the above problems,this thesis proposes an efficient and widely applicable fine-grained vehicle color recognition algorithm.Firstly,we collect a rear shooting dataset from vehicle bayonet monitoring for fine-grained vehicle color recognition.Using the prior information of the car model,we propose a labeling method combining unsupervised machine learning and manual to perform fine-grained labeling on the dataset.The labeling algorithm can eliminate the influence of light and determine the accurate color label for each image.Secondly,this thesis proposes a fine-grained classification algorithm for vehicle color based on hybrid structure.Specifically,a learnable deep convolutional neural network is used to extract feature maps from vehicle images and further converted into pixel embedding sequence,thereby enhancing the local feature expression ability of the overall network;self-attention mechanism in the Transformer is used to learn the relevance between pixels and capture the features of key areas that can reflect the color of the vehicle;combining the flag output of the different layers of the Transformer Encoder to fuse the feature information of different granularities;according to the characteristics of the proposed vehicle color dataset,we design a multi-granularity oriented hierarchical loss function for color classification.By comparing with traditional convolutional neural networks,fine-grained classification networks and color classification networks respectively,it shows the proposed algorithm has the best performance on the task of fine-grained vehicle color classification.Finally,based on the above algorithm,a prototype system for fine-grained classification of vehicle colors was developed to meet the practical application requirements of vehicle color classification.