去除霧/煙/沙及影像增強基於人類視覺系統啟發之神經模型

Abstract

惡劣環境下的低能見度是很多電腦視覺應用的主要問題，像是室外物件辨識、偵測、智能車輛、物件追蹤、監視等，效能都依賴於影像的品質。然而，因為大氣中的懸浮粒子如煙、霧、沙子、灰塵等大量的存在，使光線衰減且產生色散現象進而降低影像中場景的能見度，同時也會降低影像的對比度。這些狀況不僅會困擾和混淆人類的觀察者，還會降低那些依賴微小特徵的電腦視覺演算法之有效性與精確度，因此模擬在各種環境下所造成的視覺影響和發展出一套良好的演算法來移除與消除這些由懸浮微粒及光線衰減所造成的影響，變成不可或缺任務。 在本篇論文，首先介紹了人類視覺系統如何產生色彩視覺與探討含有煙霧圖片之模型，應用這兩項做了很多的應用，包括白天/夜晚的除霧、除煙、低光源影像增強、除沙與色彩恆常性。一開始，基於opponent based色彩恆常性演算法，進而提出了改良的版本。第二部分先探討了幾篇存在的白天/夜晚除霧方法，分別是dark channel prior、retina inspired method、glow estimate based method，說明了白天與夜晚霧霾影像的差別，提出了幾個新的白天/夜晚除霧演算法，同時，基於白天除霧演算法，達成了低光源影像增強。第三部份探討了白天/夜晚影像除煙，我們觀察到煙霧影像中，在煙霧區域的色彩會有不均勻的失真現象，我們運用dark channel的概念，成功地解決發現的問題，最終達到了白天與夜晚除煙的效果。最後的部分，我們提出了影像除沙演算法。本篇論文成功的提高惡劣環境下影像的能見度。

Poor visibility in bad weather is a major problem for many applications of computer vision such as outdoor object recognition, detection, tracking, intelligent vehicles and surveillance rely heavily on the quality of image scenes. However, bad weather conditions caused by suspending particles in the air, such as haze, sand, fog, dust, and smoke that have significant size and distribution in the participating medium. These conditions may significantly degrade the visibility of a scene due to the considerable presence of particles in the atmosphere that attenuation and scatter light. These particles suspending in air result in various degrees of attenuation, scattering and absorption the light in the atmosphere. This effect may significantly reduce the contrast, limit the visibility and faded the colors of the daytime scenes and nighttime scenes, resulting in a severely degraded image. It attenuates the signal of the viewed scene. Then, impacts negatively on the accuracy of many applications of computer vision. Therefore, enhancing visibility is an inevitable task. In this thesis, we introduce about how human visual system (HVS) and haze image model can be applied in many fields such as daytime/nighttime image dahazing, color constancy, low-light enhancement, daytime/nighttime image desmoking and sand removal. The first part of the thesis is to introduce the effect of human visual system and haze image model. Then, apply these models to color constancy algorithm. The second part is about some important existing daytime/nighttime dehazing algorithm based on haze image model and HVS. We observe some differences between nighttime and daytime hazy images. First, atmospheric light in nighttime hazy images suffer from non-uniform illumination and glowing effect. Second, nighttime hazy images have low illumination and some details get lost under insufficient illuminance. Third, visible lights sources with varying colors will cause an obviously color shift in the image. We propose some new daytime/nighttime dehazing models to solve these three problems and use two daytime dahazing methods to achieve low-light enhancement algorithm. In the third part, we observe that the unbalanced particle density distributed in each RGB color channel make the smoke region of smoky images suffer from hue distortion. Moreover, the smoke region is non-homogeneous which means that the concentration of the smoke is not approximate the same in the entire scene. We propose some novel daytime/nighttime smoke removal models based on haze image model to successfully address these problems. In the last part, we propose the sand removal algorithm to remove the sandstorm in the images.