Semantic Segmentation and Object Detection Based On Active Contour Model and Fuzzy Clustering.

摘要

We propose a novel method for image segmentation and object detection. The proposed strategy is based on two major steps. The first step corresponds to image segmentation which is based on Active Contour Model (ACM) algorithm. The gradient stopping function has been widely used in most ACMs as an edge indicator. Because of the gradient high sensitivity to texture and noise, other stopping functions, such as polarity, have been proposed with some limited success. Unfortunately, most of these proposed stopping functions, including gradient and polarity, fail to detect objects effectively in many circumstances. On the other hand, depth information, if available, could provide the better clue for object detection. The proposed method takes the advantage of the existing contour models by using the depth clue, from either Kinect sensor or stereo vision algorithm, instead of two-dimensional clues, in the model stopping function. However, even with depth clue, it is still difficult to accurately detect a salient object when it is located at similar depths of others. Indeed, based on specific image data or genre of the image, the best candidate for a stopping term could be either a single feature such as gradient, polarity or depth, or a combination of them. So the proposed ACM is based an automatic selection of best candidate features among gradient, polarity and depth, coupled with a combination of them by Kernel Support Vector Machine (KSVM). Although existing techniques, such as the ones based on ACM perform quite well in the single-object case and non-noisy environment, these techniques fail when the scene consists of multiple occluding objects, with possibly similar colors. Thus, the next step corresponds to the identification of salient and occluded objects based on Fuzzy C-Mean (FCM) algorithm. In this latter step, the depth is included as an important clue that allows us to estimate the cluster number and to make the clustering process more robust. In particular, occlusions are easily handled this way, and the objects are properly segmented and identified. The experiments, carried out on real images, have shown the success and effectiveness of our proposed method to detect the salient objects.