Adaptive Region Clustering in LDA Framework for Image Segmentation

Image segmentation based on low-level features has been studied for many years. However, because of the semantic gap issue, it is difficult to have more breakthroughs based on low-level features. LDA is a powerful tool to model co-occurrence relationships between words and thus is used to catch the semantic connections between low-level visual features. In image segmentation, the codebook is built from the visual features and topics are trained with LDA model. And the topic distributions yield important guidance for segmentation. However, in previous papers, researchers used the topic with the highest probability to merge the regions. It ignored the statistics nature of the topic distribution. And, the segmentation result will be greatly impact by the codebook size, the topic number and cluster number. To address these challenges, this paper proposes a new image segmentation algorithm based on LDA framework: an adaptive region clustering approach based on EM. We build the cookbook from the color, texture and SIFT features and perform the LDA training using Gibbs Sampling for topics. Then the adaptive region clustering with EM is invented to merge the regions based on topic distribution. The clustering number is self-identified according to Minimum Description Length (MDL) principle. And an image is represented as a Gaussian Mixture Model (GMM) with objects corresponding to Gaussian mixture components. The final segmentation could be achieved after the region clustering and adjacent check. We implemented the new algorithm and conducted experiments to validate the region clustering approach and segmentation performance. And the results show great effectiveness of this new algorithm.