電腦視覺在近幾年高度發展，在物件偵測方面，YOLO、R-CNN 已經有不錯表現。如果輸入影像因拍攝時隔著玻璃，受到反射影干擾，這些物件偵測演算法會受到影響，導致定位失準或甚至遺漏。因此本論文希望提出一個基於生成對抗網路 (Generative Adversarial Network) 對單一影像進行反射影去除的演算法，來降低其影響，使得物件偵測效果能夠有所改善。 本論文以行車紀錄器影像為主要應用情境，偵測包含汽車、機車、卡車、 公車、腳踏車五種常見的交通工具。由於行車紀錄器常安裝於汽車擋風玻璃後，影像有機會受到玻璃上的反射影影響，進而使物件偵測效果變差，從而影響駕駛輔助系統。所以，本論文先訓練 Mask R-CNN，並用以偵測反射影及其位置，再訓練一個生成對抗網路，使其可以利用反射影後方的特徵進行還原，達到反射影去除的效果，藉此進一步改善物件偵測的準確度。;Computer vision has been highly developed in recent years. In terms of object detection, YOLO and R-CNN have already performed well. However, if the input image is disturbed by reflections due to the separation of the glass during shooting, these object detection algorithms will be affected, resulting in misalignment or even omission. Therefore, this thesis tries to propose an algorithm based on Generative Adversarial Network (GAN) to remove the reflection of the single image to reduce its impact and improve the object detection effect. This thesis uses the dashcam image as the main application scenario to detect five common vehicles including cars, motorbikes, trucks, buses, and bicycles. Since the dashcam is usually installed behind the windshield of the car, the image may be affected by the reflection on it, which may make the accuracy of the object detection worse, thereby affecting the driving assistance system. Therefore, this thesis first trains a Mask R-CNN to detect the mask of the reflection, and then trains the GAN so that it can use the features behind the reflection to restore and achieve the effect of reflection removal. Further raises mAP of object detection.
