影像缺陷偵測一直是影像處理在研究的問題。如何定義影像缺陷是傳統影像處理演算法需要面對的問題。近幾年深度學習在影像處理與電腦視覺的領域上都有著亮眼的突破,許多深度學習模型在移除影像缺陷的資料集上都有不錯的結果,像是去雜訊、去模糊、亮度調整等。但不是所有影像經過這些深度學習模型處理後都會變得更好看,有些原本沒有缺陷的影像在經過模型處理後,顏色反而會變得詭異或出現不自然的雜訊。為了能使用既有的移除影像缺陷方法作為影像缺陷偵測的依據,本文提出一個新的架構,藉由整合影像增強與影像品質評估,篩選掉影像增強後效果不好的結果,達成影像缺陷的自動偵測。在與曝光相關的資料集上實驗後,我們發現影像品質評估有能力找到曝光較好的圖片。我們接著結合兩個不同的影像曝光增強方法,讓影像品質評估判斷何者具有更好的品質,最後列出影像品質評估分數差距最大的前幾名,確認這些圖片是否具有曝光的問題。
Image defect detection has been a research topic in image processing for a long time. How to define image defects is a problem that traditional image processing algorithms need to face. Deep learning has made outstanding breakthroughs in image processing and computer vision in recent years. Many deep learning models for removing image defects, such as denoising, deblurring, and brightness modification, have shown promising results in many datasets. However, these deep learning models will not improve the appearance of all images. Some images will have unnatural color or noise after being processed with deep learning models. In order to use the existing image defect removal methods as the basis for image defect detection, we propose a new framework that integrates image enhancement methods and image quality assessment to filter out the poor enhancement results and achieve the automatic image defect detection. We discovered that image quality assessment is capable of identifying better exposed images after experiments on exposure datasets. There are two image exposure enhancement approaches in our experiments. One is gamma correction, and the other one is “Learning Multi-Scale Photo Exposure Correction” model. We used the two image exposure enhancement methods to enable the image quality assessment to identify which one has better quality. Finally, we listed images with the highest image quality assessment score difference to confirm whether these images have exposure problem or not.