”Combination of Experts” Approach to Image Boundary Detection

Boundary detection in two-dimensional images is an important problem in computer vision. There are a wide variety of algorithms to accomplish this task, but none have come close to human proficiency. We explore a variety of ways to use machine learning algorithms to combine existing boundary detection algorithms with the goal of exceeding the performance of any particular algorithm. We present methods using Adaboost and linear regression and show favorable results produced by them. Boundary detection is a classification problem which entails labeling a subset of pixels in an image as part of an edge that separates two objects. However, there is no clear definition for which objects should be separated. The most common approach is to attempt to find a set of edges which a human being would consider reasonable. Nonetheless, for any given image, there may be disagreement among human beings as to what set of edges best captures the image. Additionally, while some boundary detection algorithms impose hard boundaries, others instead provide probabilities that each pixel in an image is part of an edge. Boundary detection, an active research area within the artificial intelligence field of computer vision, has a variety of applications to higher-level vision tasks. Many object-recognition algorithms use image boundaries as inputs. Furthermore, boundary detection algorithms help show the precise orientation of an object in space, which is useful for robotic manipulation tasks. This approach to boundary detection is particularly interesting to us, since we are working on an MRF-based image segmentation algorithm that takes the output from a boundary detection algorithm as its input. Thus, this project could provide better inputs for image segmentation. Alternatively, it may allow us to combine the image segmentation algorithm’s boundary output with other boundary detection algorithms’ outputs for even better results.