License Plate Super-Resolution by Edge enhanced neural network

Due to the complex structure of Chinese characters, which are composed of multiple strokes with varying shapes and lengths, existing super-resolution (SR) methods still suffer from the issue of Chinese characters distortion when applied to license plate SR tasks. To address this problem, we propose a novel edge information correlate SR network (ECSR) with an embedded edge information correlate block. The proposed network decouples the edge information of license plates (LPs) into two major directions (horizontal vertical and diagonal) to take advantage of the edge information of LR images and enhanced the reconstructed HR images. As relevant datasets for LP SR are currently lacking and text details tend to be lost during LP image acquisition, we introduce a multi-scene degraded (noise, low light, motion blur, high light, lowlight + noise) LP SR dataset, called CLP22359. Reconstruction and LP recognition experiments on CLP22359 demonstrate the superiority of the proposed method over state-of-the-art methods. The datasets we created for this work are available at https://github.com/1312677659/CLP22359.