車牌文字品質評估

Abstract

在行車紀錄、道路或停車場監控的應用中，所拍攝的車牌影像並不需要有極高的畫面品質，但必須足夠清晰，使得車牌中的文字能被正確地辨識。因此，本篇論文提出一種車牌文字品質評估方法，對於壓縮後的車牌，評估車牌上的文字能被辨認出的程度。因此，有別於傳統的品質評估方法，如基於像素差異(例如：PSNR)或結構特徵(例如：SSIM)，本論文所使用的是基於文字的特性，作為品質評估的依據，提出一種以語意為主（semantic-based）的品質評估演算法。 本篇論文結合了影像品質評估及車牌辨識，對於車牌壓縮圖片中，透過文字辨識的演算法先行計算圖片中文字的位置，再基於文字辨識的原理，利用雜訊對文字影響的程度計算雜訊密度較高的部分及模糊程度來評估出車牌的品質。在未來應中用中能夠以分數高低判斷是否能用較少的bits數傳遞圖片，在相同能讓電腦辨識出來的情況下有效減少傳輸圖片的bits數，提升效能。我們所提出來的影像文字品質評估方法(Image Semantic Quality Assessment, ISQA)，和一些熟知的影像品質評估方法等作比較，在Spearman及Kendall的相關係數上有顯著的提升。

In this thesis, we proposed an image semantic quality assessment method for car-plate images. The purpose of our method is to evaluate whether the characters in a car-plate image can be recognized or not after compressed. To this end, we considered that the compressed car-plate image has to be calculated from semantic-related features, rather than pixel-wised (ex. PSNR) features or structure-wised features (ex. SSIM). The proposed image semantic quality assessment (ISQA) method is based on car-plate recognition (CPR) techniques. By considering text locations, our algorithm combines high density detail blocks with blur to calculate the quality score for compressed car-plate images. The result can be applied to judge whether lower bitrates can be used in image compression to achieve the same recognition results, and hence improve the image coding efficiency. The proposed image semantic quality assessment method (ISQA) has been compared to some related image quality assessment metrics, and the result shows that both Spearman and Kendall correlation coefficients can be improved significantly.