This paper presents a scene classification method based on probabilistic Latent Semantic Analysis (pLSA) and a k-nearest neighbour classifier. The goal is to discover objects in scene images in an unsupervised manner and use this object distribution for scene classification. pLSA, originally developed for text analysis, is applied here to a bag-of-visual-words representation of images. The scene classification is performed using a k-nearest neighbour classifier. The paper investigates the classification performance under changes in the visual vocabulary and number of latent topics. A novel vocabulary is developed using colour SIFT descriptors. The method is compared to supervised approaches (Vogel & Schiele, Oliva & Torralba) and a semi-supervised approach (Fei Fei & Perona). The combination of unsupervised pLSA followed by supervised nearest neighbour classification achieves superior results. The paper introduces a new classification algorithm based on pLSA followed by a nearest neighbour classifier. The pLSA model is applied to images represented by the frequency of "visual words". The formation and performance of this visual vocabulary are investigated, comparing sparse and dense feature descriptors. The approach is inspired by previous works using pLSA on sparse features, dense SIFT features for pedestrian detection, and semi-supervised LDA for scene classification. The paper extends these works by developing new features and improving the classification algorithm. The paper compares its classification performance to three previous methods using the authors' own databases. Previous works used varying levels of supervision in training, while this paper uses an unsupervised object discovery approach. The results show superior performance in all cases. Section 2 describes the pLSA model, which is used to discover topics (object categories) in images. Section 3 describes the classification algorithm based on applying pLSA to images. Section 4 describes the features used to form the visual vocabulary and the principal parameters investigated. Sections 5 and 6 describe the datasets and experimental evaluation.This paper presents a scene classification method based on probabilistic Latent Semantic Analysis (pLSA) and a k-nearest neighbour classifier. The goal is to discover objects in scene images in an unsupervised manner and use this object distribution for scene classification. pLSA, originally developed for text analysis, is applied here to a bag-of-visual-words representation of images. The scene classification is performed using a k-nearest neighbour classifier. The paper investigates the classification performance under changes in the visual vocabulary and number of latent topics. A novel vocabulary is developed using colour SIFT descriptors. The method is compared to supervised approaches (Vogel & Schiele, Oliva & Torralba) and a semi-supervised approach (Fei Fei & Perona). The combination of unsupervised pLSA followed by supervised nearest neighbour classification achieves superior results. The paper introduces a new classification algorithm based on pLSA followed by a nearest neighbour classifier. The pLSA model is applied to images represented by the frequency of "visual words". The formation and performance of this visual vocabulary are investigated, comparing sparse and dense feature descriptors. The approach is inspired by previous works using pLSA on sparse features, dense SIFT features for pedestrian detection, and semi-supervised LDA for scene classification. The paper extends these works by developing new features and improving the classification algorithm. The paper compares its classification performance to three previous methods using the authors' own databases. Previous works used varying levels of supervision in training, while this paper uses an unsupervised object discovery approach. The results show superior performance in all cases. Section 2 describes the pLSA model, which is used to discover topics (object categories) in images. Section 3 describes the classification algorithm based on applying pLSA to images. Section 4 describes the features used to form the visual vocabulary and the principal parameters investigated. Sections 5 and 6 describe the datasets and experimental evaluation.