In recent years, scene semantic recognition has become the most exciting and fastest growing research topic. Lots of scene semantic analysis methods thus have been proposed for better scene content interpretation. By using latent Dirichlet allocation (LDA) to deduce the effective topic features, the accuracy of image semantic recognition has been significantly improved. Besides, the method of extracting deep features by layer-by-layer iterative computation using convolutional neural networks (CNNs) has achieved great success in image recognition. The paper proposes a method called DF-LDA, which is a hybrid supervised-unsupervised method combined CNNs with LDA to extract image topics. This method uses CNNs to explore visual features that are more suitable for scene images, and group the features of salient semantics into visual topics through topic models. In contrast to the LDA as a tool for simply extracting image semantics, our approach achieves better performance on three datasets that contain various scene categories.