In the past few years, significant progresses have been made in scene segmentation and semantic labeling by integrating informative context information with random field models. However, many methods often suffer the computational challenges due to training of the random field models. In this work, we present a fast approach to obtain semantic scene segmentation with high precision, which captures the local, regional and global information of images. The approach works in three steps as follows: First, an intermediate space with lowdimension semantic "topic" representation for image patches is introduced, by relying on the supervised Probabilistic Latent Semantic Analysis. Secondly, a concatenated pattern is taken to combine the vectors of posterior topic probabilities on different feature channels and to incorporate them into a conditional random field model. Finally, a fast max-margin training method is employed to learn the thousands of parameters quickly and to avoid approximation of the partition function in maximum likelihood learning. The comparison experiments on four multiclass image segmentation databases show that our approach can achieve more precise segmentation results and work faster than that of the state-of-the-art approaches.