疊代推進生成對抗網路用於陰影去除
No Thumbnail Available
Date
2022
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
隨著科技的高速發展，深度學習在工業、軍事、民生科技處處都有大量的應用，現今運用在影像處理上的深度學習技術不斷進步，影像的去除如影像除霧、去反光、去陰影等都是電腦視覺領域中具挑戰性的任務。本論文研究目的為針對影像陰影去除提出了迭代推進生成對抗網路，首先我們輸入陰影圖藉由兩個生成器網路分別生成出無陰影的圖及殘差陰影圖，將兩者合成得到陰影圖，與輸入進行比對，最後將合成的圖再次輸入至網路重複上述步驟直到收斂，透過迭代推進的方式提升陰影移除的效果。此外為了使結果更加優異，我們的生成器網路加入了注意力機制，讓模型更專注於影子的部分，以及長短期記憶，使我們在長序列訓練過程中有更好的表現，最後是修復網路，以進一步改善生成的結果。我們與傳統方法以及近年來基於深度學習所提出的陰影去除方法比較，實驗結果表明本論文所提出的迭代推進方法有更優異的結果。
With the rapid development of technology. Deep learning used in image processing is constantly advancing. Image removal such as image haze removal, reflection removal and shadow removal are all challenging tasks in the field of computer vision. The purpose of this paper is to propose an iterative advance generative adversarial network for image shadow removal. First, we input the shadow image through two generator networks to produce shadow-free image and residual shadow. The outputs of the two networks are combined to compare the input image. Through an iteratively advance manner, the effect of shadow removal has a great improvement. In order to make the results more excellent. The generators networks contain attention mechanism so models can more focus on the shadow portion and the Long Short-Term Memory to improve training through long sequence training. Then an inpainting network is applied to further improve the results.
With the rapid development of technology. Deep learning used in image processing is constantly advancing. Image removal such as image haze removal, reflection removal and shadow removal are all challenging tasks in the field of computer vision. The purpose of this paper is to propose an iterative advance generative adversarial network for image shadow removal. First, we input the shadow image through two generator networks to produce shadow-free image and residual shadow. The outputs of the two networks are combined to compare the input image. Through an iteratively advance manner, the effect of shadow removal has a great improvement. In order to make the results more excellent. The generators networks contain attention mechanism so models can more focus on the shadow portion and the Long Short-Term Memory to improve training through long sequence training. Then an inpainting network is applied to further improve the results.
Description
Keywords
陰影去除, 生成對抗網路, 卷積神經網路, 深度學習, shadow removal, generative adversarial network, convolution neural network, deep learning