一种改进多标签分类结构的实时车牌识别模型

An improved multi-label classification structure for real-time license plate recognition model

本文提出一种改进的多标签分类结构的实时车牌识别模型,通过压缩-激活注意力机制在通道方向关注车牌图像中的字符,并且融合网络的深浅层特征,在提取语义的同时又能兼顾到字符定位,提高字符定位精度,同时采用深度可分离卷积取代传统卷积,减少计算量,提高车牌识别速度。最后,在CCPD数据集中选取最能体现光照强度、车牌倾斜和车牌模糊等复杂场景的3个子数据集进行测试实验,测试结果达到94.0%的平均准确率和395.6 FPS的识别速度,这表明本文提出的模型能够在复杂场景下达到快速识别车牌的效果。

This paper proposes an improved multi-label classification structure for real-time license plate recognition model designed using deep learning technology.The model utilizes the Squeeze and Excitation attention mechanism to focus on license plate characters in the channel direction and combines features from the network's deep and shallow layers.This approach realizes character localization and semantic feature extraction,improving character positioning accuracy.Additionally,the model adopts Depthwise Separable Convolution in place of traditional convolution to reduce computational complexity and improve license plate recognition speed.Finally,the proposed model is tested on three subsets of the CCPD dataset that demonstrate complex scenarios,such as varying illumination,plate inclination,and blurriness.The test results show an average accuracy of 94.0%and a recognition speed of 395.6 FPS,indicating the proposed model's effectiveness in quickly identifying license plates in complex environments.