Contents I. Introduction
Truck recognition is one of the key important methods of truck attribute recognition, which is widely used in the fields of vehicle recognition in expressway scene, non-stop charging, truck violation inspection and so on. Truck color information is one of the key information of vehicle besides truck license plate information. By truck color recognition, the weakness of vehicle recognition based on single vehicle attribute could be made up, which could play an important role in combating false license plate. However, because of the influence of factors such as long vehicle driving time, complex road scene and illuminations changes, it is difficult and challenging to realize accurate recognition of truck color through truck image. Therefore, how to realize the accurate recognition of truck color through truck image has become an urgent problem to be solved in the field of intelligent transportation system. In 2017, Xue et al. [1] adopted a vehicle image fusion processing method including histogram equalization method, local contrast enhancement and homomorphic filtering method to process the vehicle image and improve the accuracy of vehicle color recognition. In 2017, Aarathi et al. [2] proposed a method to solve the problem that vehicle images taken from roads or hill areas could not be effectively recognized because of haze, which mainly adopts the dark channel prior technique and CNN to remove the haze and learn feature respectively. In 2018, Kim et al. [3] proposed a vehicle color classification method based on representative color region extraction and convolutional neural (CNN), which randomly selects points from the probability map of representative color region produced by Harris corner detection method to generate an input image for CNN model training. In 2018, Zhang et al. [4] proposed a vehicle color recognition method based on a lightweight CNN which contains three convolutional layers, a global pooling layer and a fully connection layer. Compared with traditional vehicle color recognition method, this method could reduce the dimension of feature vector and the memory occupation of the model, at the same time improve model accuracy slightly. In 2019, Sun et al. [5] integrated the trained vehicle brand recognition network and vehicle color recognition network based on the training mode of multi-task learning, and constructed a vehicle multi-attribute recognition model. In 2019, Zhang et al. [6] built a vehicle color recognition network model based on deep convolution neural network by adjusting structure and parameters of Deep- Vgg-16 model. In 2020, Fu et al. [7] put forward a multi-scale comprehensive feature fusion convolutional neural network (MCFF-CNN) based on residual learning to solve the problem of automatic vehicle color conditions, which first extracts the dark color features of vehicles through MCFF-CNN network, and then through support vector machine (SVM) classifier to obtain the final color recognition results. In 2021, Tariq et al. [8] proposed a vehicle detection and vehicle color classification method based on Faster R-CNN to solve the problem that the commonly used recognition methods rely heavily on hand-made features. In 2021, Awang et al. [9] studied at the impact of different schemes of color images on the performance of vehicle type recognition system, and compared the performance of vehicle feature extraction models under YCrCb and RGB color schemes. In 2021, Hu et al. [10] built a vehicle color database contains 24 vehicle colors and proposed a Smooth Modulated Neural Network with Multi-layer Feature Representation (SMNN-MFR) to solve the problem of long tail distribution in the dataset.
