Object Recognition Based On ML_pLSA Model And Bag-of-Feature Algorithm

Object recognition is one of the hottest topics in computer vision field. It has been drawn more and more attention, and has act as an important role in satellite remote sensing image analysis, medical image analysis, face recognition and vehicle license plate recognition, etc. However, object recognition technique still suffers from several challenges such as object rotation change, occlusion, object scale and illumination variations.This paper presents a new object recognition method based on Multi-Level-probabilistic Latent Semantic Analysis (ML-pLSA) object recognition and Bag-of-features algorithm. Traditional segmentation based recognition methods segment the image first and then try to find the segments that contain the object. On the contrary, our proposed method joins object segmentation and recognition together, and makes the two parts promote and help each other, to avoid the condition that segmentation based recognition methods rely too much on the quality of image segments. The process of our method is as follows. Firstly, multiple segmentations at different levels are computed for each image, and then object classes on each segment region are estimated by using pLSA and bag-of-words. The final results are obtained by fusing estimation results at multiple levels. Compared to traditional object recognition algorithms, our method is more robust and applicable, and works well for unsupervised learning.This paper introduces the probabilistic Latent Semantic Analysis model which works successfully in text retrieval field into computer vision field, and proposes ML_pLSA (Multi-Level-probabilistic Latent Semantic Analysis) model. This model utilize the capability that pLSA model can deal with the synonym and polyseme problem to tackle the inter-class and intra-class variation of the images. The probability distribution, conditional distribution and all the parameters of the model are demonstrate clearly in the paper. The experiment results on Graz-02 dataset demonstrate that the proposed method performs better than traditional object recognition methods in both accuracy and robustness.