改进注意力机制实现车牌图像清晰化

Realizing License Plate Image Enhancement by Improving Attention Mechanism

为避免低分辨率模糊图像对识别车牌的影响,提出基于残差网络的空间和通道双重注意力网络(RSCAN)来恢复模糊车牌图像。RSCAN在Residual Channel Attention(RCAN)网络基础上添加空间注意力机制来捕获更多空间信息,引入新的通道注意力机制加强采集图片通道信息能力,以及改进损失函数,加入梯度损失函数改善重建车牌图片的视觉效果,通过自制的车牌数据集来训练和测试网络。使用测试车牌图片将RSCAN网络与常规处理图片、EDSR(Super-resolution)网络、RCAN网络进行对比实验,通过峰值信噪比(PSNR)和结构相似性(SSIM)进行模型评估,分别为31.284/0.862。验证结果表明,设计的RSCAN网络处理车牌图像和其他方法相比取得很好的效果。

In order to avoid the impact of low-resolution blurred images on the recognition of license plates,a spatial and channel dual attention network(RSCAN)based on residual network to recover blurred license plate images is proposed.RSCAN adds a spatial attention mechanism to the Residual Channel Attention(RCAN)network to capture more spatial information,introduces a new channel attention mechanism to enhance the ability to collect picture channel information,and improves the loss function.Adding a gradient loss function to improve the visual effects of reconstruction of license plate images,train and test the network through a self-made license plate data set.The RSCAN network is compared with conventional processed images,EDSR(Super-resolution)networks,and RCAN networks,using test license plate images for comparison experiments,and model evaluation is conducted through peak signal-to-noise ratio(PSNR)and structural similarity(SSIM),which are 31.284/0.862 respectively.The experimental results of the system test experiment show that the license plate image processed by the RSCAN network achieves better results compared with other methods.