Category-Consistent Deep Network Learning for Accurate Vehicle Logo Recognition
Neurocomputing, 2021, 463: 623-636.
Wanglong Lu, Hanli Zhao, Qi He, Hui Huang, Xiaogang Jin
OurOverall pipeline of the proposed category-consistent deep network learning framework. We take images in the same category to generate category consistent masks by using a pre-trained VGG-19 model and deep descriptor transformation algorithm. Then, each pair of image and mask is fed into the proposed category-consistent deep network for training. The feature map extracted with the VLF-net module is further fed into the classification module and CCML module. After end-to-end training, these modules can jointly optimize the representation learning of the backbone network.
Abstract
Vehicle logo
recognition (VLR) is essential in intelligent transportation systems.
Although many VLR algorithms have been proposed, efficient and accurate VLR
remains challenging in machine vision. Many VLR algorithms explicitly detect
the coarse region of the vehicle logo either by offsetting the detected
location of the license plate or by training on numerous images with manual
bounding-box annotations. However, the results of license plate detection
can significantly influence the VLR accuracy, whereas bounding-box
annotations are considerably labor-intensive. Thus, we propose a novel
category-consistent deep network learning framework for accurate VLR. A
convolutional-neural-network-based vehicle logo feature extraction model is
proposed to extract deep features by considering both high- and low-level
features in an image. Moreover, a novel category-consistent mask learning
module is proposed to help the framework to focus on category-consistent
regions without relying on license plate detection or manual box
annotations. The deep network is trained and optimized iteratively with the
objective function incorporating classification loss and
category-consistency loss. Extensive experimental evaluations and
comparisons on the publicly available HFUT, XMU, CompCars, and VLD-45
datasets demonstrate the feasibility and superiority of the proposed
algorithm.