Image Segmentation Based on Weakly Supervised MKL on Mixed Visual Features

Weakly supervised learning has outstanding ability to solve classification tasks, and multiformity middle-level visual features provide more abundant discriminant information for meaningful regions. In this paper, we study the integration of the middle-level visual features including homogeneity of superpixels, region objectness and texture map for segmentation. Then, three kernels are exploited to map visual features to high-dimensional space. A few labeled pixels are chosen for training support vector machines(SVMs) in a single image with hybrid kernels. On this basis, the remaining pixels are labeled with classified results of SVMs and refined the segmentation results by merging pre-segments of mean-shift. We perform sufficient experiments on Berkeley datasets and compared them with several excellent segmentation algorithms. Extensive experimental results of the proposed method show superior segmentation performance and expanded tests on PASCAL VOC datasets further validate the effectiveness of the algorithm.