The Vehicle License Plate Color Detection Algorithm Optimized Based on a Parameter-Free Attention Mechanism

Chenpeng Shi,Jianrong Xu,Chao Zhang,Sen Wang,Yuming Liu,Xiaoyong Zhang

DOI: https://doi.org/10.1145/3677182.3677202

2024-01-01

Abstract:In response to the challenges posed by high false-positive rates and considerable inference costs in license plate detection for moving vehicles within the context of road surveillance, we propose a novel algorithm for license plate detection and color recognition. This algorithm is specifically designed to address the requirements of small target recognition in roadside monitoring scenarios and is optimized based on a parameter-free attention mechanism. Built upon the YOLOv5s architecture, the algorithm optimizes the model structure by retaining critical components of the original design, utilizing smaller convolutional kernels, and reducing the number of channels. Additionally, it introduces an innovative three-dimensional attention mechanism module in the detection network, resulting in enhanced inference performance and recognition accuracy at an equivalent parameter scale. To further enhance the optimization efficiency and convergence effectiveness of model training, we implemented prior anchor boxes tailored for roadside license plate detection. Experimental results demonstrate that, without compromising the model's recognition performance, the optimized model has a parameter count of only 0.22 M, a model storage size of merely 0.6 MB, and an inference cost reduced to 1/10th of the original optimization.