使用卷積神經網路架構進行單張影像煙霧移除

Abstract

煙霧常常會使得影像中能見度以及物件辨識準確度變低，而基於大氣模型進行除霧演的算法已經是一個被廣為研究的領域。但是，因為煙在影像上有著高度空間上的特性，例如煙在影像上的位置，以及，在鄰近的像素中比起霧有著更高範圍的濃度變化，使得除煙演算法並還沒有像除霧演算法一樣被廣泛的研究。也因上述的原因，直接使用除霧演算法為不切實際的做法，這些除霧演算法往往會造成影像色失真，尤其在無煙區域，進而使能見度及物件辨識的準確度並沒有有效提升。而為了解決這些問題，我們提出了一套基於卷積神經網路所組成的架構，這架構中包含了兩個子模型–煙區偵測網路(SRD-Net)和除煙網路(Desmoke-Net)來進行除煙。煙區偵測網路會先進行煙區的偵測，再者，使用除煙網路對所偵測的煙區來預估大氣模型的煙霧濃度圖，得到了煙霧濃度圖便能以大氣模型進行圖片能見度的修復。這套是第一個只需單張影像便能進行的除煙演算法，而且還保證了在無煙區域並不會有任何顏色上的改變，進而可以避免影像色失真，再者，其除煙的效果也比現行上的除霧演算法有著更為顯著的效果。

Smoke and haze usually deteriorate visibility of images and accuracy of object detection. Atmospheric model has been a popular approach to remove haze with single image. However, smoke removal, which possesses stronger spatial characteristic on image, has not been well studied yet due to difficulty of smoke positioning and high range intensity changes between neighbor pixels. Thus, it is impractical to directly apply existing haze removal algorithms on smoke removal. Moreover, color distortion is a common defect caused by haze removal algorithms. To solve problems mentioned above, a novel convolutional neural network architecture including two sub-models is proposed to recover smoky images: Smoke Region Detection Net (SRD-Net) is responsible for identifying smoky pixels. Desmoke-Net estimates transmission map, which is a key factor of atmospheric model, of smoky pixels. Images can be restored via atmospheric model and its estimated transmission map. The proposed architecture is the first smoke removal algorithm requiring single image only. Values of non-smoky region pixels also remain identical to avoid color distortion. Lastly, performance is better than existing haze removal algorithms on smoke removal.