Shadow Detection In Complex Scene

Shadow is everywhere but ignored easily by us. However it’s meaningful for us to study shadow. Shadow detection can help us to understand the images, such as judging the strength and identifying the direction of the light. We can also estimate the physical properties of the object by shadow detecting. However several high-level computer vision tasks such as object detection, recognition and tracking depends on the quality of the input images. But in the real world, shadows may cause serious interference of these algorithms. Human beings can recognize the shadow regions immediately, while shadow detection is very challenging in computer vision task. It’s very important for us to detect the shadows of the region that we interested, especially detecting the shadow cast of the foreground objects. Recent approaches have mainly used illumination intensity variants, color constancy, texture invariants which can fail severely when the qualities of the images are not very good or the images are gray. For the shortages of using artificial design image features, we labeled the image datasets, and used the PC ANet to extract the features of the shadow regions, after that we trained the SVM classifier to decide whether an image edge is a shadow on the edge. Finally we applied Conditional Random Filed to optimize the contours of the shadows.