Object recognition based on probabilistic Latent Semantic Analysis (pLSA) has shown excellent performance, but it is sensitive to background clutter. In this paper, we propose a novel framework called AM-pLSA, which combines pLSA with visual attention model, to learn object classes from unlabeled images with cluttered background. We firstly detect salient regions and non-salient regions in an image using visual attention model, assuming that objects to be learned are in salient regions. By this way, we can segment interested objects from images, reducing the influence of background clutter. Then, we model each region as a visual word histogram, and learn objects classes from these regions using pLSA. Experimental results showed that AM-pLSA evidently outperformed pLSA, and was more robust to background clutter.