應用機器學習技術於有缺陷之照片的分類法

Abstract

隨著影像擷取裝置逐漸普及化，大量的照片被產生出來。然而，並非所有的照片都有好的 品質。在此論文中，我們提出了兩個影像偵測的方法，分別是模糊偵測與曝光偵測。 模糊偵測利用支援向量機器(support vector machines)估算一張影像的模糊程度，同時 判斷此影像是全域性的模糊抑或是區域性的。對於全域性的模糊影像，我們計算其點光源 擴散函數(point spread function) 並將其分類成相機搖晃或是失焦(out of focus)所造成的模 糊。對於區域性的模糊影像，我們利用影像分割的方法找出模糊的區塊，並利用點光源擴 散函數的計算將其分類成景深(depth of field)或是移動物體我產生的模糊。 和模糊問題一樣，曝光也是一個造成影像品質下降的常見的理由。結合了感興趣區 域(region of interest)偵測技術，我們可以判斷一張影像的前景/背景是否曝光過度/不足。 本論文所使用方法的優點是所有過程皆是自動進行，因此使用者可以簡單的找到他們 真正想要的影像。

Photos are massively produced while digital image capturing devices are becoming popular, however, not every photo has good quality. In this thesis, two image detectors are proposed: blur detector and exposure detector. Blur detector uses support vector machines to estimate the blur extent of an image and distinguishes the difference between locally blurred image and globally blurred image. For globally blurred image, we estimate the point spread function (PSF) and the image is classified to camera shake or out of focus. For locally blurred image, we find the blurred regions using segmentation method, and the PSF estimation on the blurred region can sort out the image with depth of field or moving object. Exposure problem, as well as blur, is another familiar cause of defective images. Combining with (ROI) estimation technique, we can tell if the background/foreground of the image is over/under exposured. The advantage of our framework is that the processes are automatic, so the users can easily find the images they want by these hints. v