Research On Commodity Image Classification

With the development of Internet, e-commerce has been developed rapidly. As the main information carrier, commodity image has become an important medium of selective purchase. Commodity image classification has been an important application branch of image classification, providing with strong sup-port of commodity retrieval, placement strategy formulation and intelligent rec-ommend. In recent years, SRC (Sparse Representation based Classification) has made great progress in image classification and has become the research hotspot in the field of image processing, computer vision and pattern recognition.In this dissertation, commodity image classification based on sparse rep-resentation is studied from two aspects of theory and application. In theory, sparse dictionary learning problem and sparse coding optimization problem are mainly focused; while the application research is carried out from symbiosis and complementarity of image and text in commodity image classification and temporal context correlation in the user browsing process. We improve the ex-isting algorithms, and verify their performance by simulation experiments.The main contributions and innovations of this dissertation are as follows:1) According to the characteristics of structured and compact of sparse dictionary, USS-DL (Unsupervised Structured Sparse Dictionary Learning) al-gorithm is proposed. It initializes the dictionary with K-means algorithm to im-prove its compactness, and constrains the objective function with sub-dictionary relevancy and classification accuracy to enhance its compactness and structural-ization in the process of dictionary learning. The experimental results on the databases of commodity image classification, face recognition and handwrit-ing recognition show that USS-DL algorithm can reduce the distance inside the class, increase the distance between classes, and improve classification perfor-mance for different types of images effectively.2) According to the characteristic of block sparseness of sparse represen-tation, BSCMN (Block Sparse Coding based l2,0 Mixed Norm) algorithm is proposed. It imposes local and global constraints on l2 and l0 norm separately for the characteristics of block structure and sparseness. The experimental re-sults on the databases of commodity image classification, face recognition and handwriting recognition show that BSCMN algorithm can improve discrimina-tion of sparse representation, and can achieve desired classification effect for different types of images.3) According to symbiosis and complementarity of commodity description text and its image, MM-IT (MultiModal based on Image and Text for classifica-tion) algorithm is proposed. It calculates the confidences of image and text clas-sification respectively using the proposed BSCMN and WB-KNN (Weighted Bayes K-Nearest Neighbor) algorithm, obtains the fusion confidence by dimen-sionless processing of two different modals, and classifies it with the proposed fusion classifier. The experimental results on commodity image classification database show that MM-IT algorithm can achieve desired classification results using text information to enhance visual information.4) According to time context semantic relation between commodity im-ages, CSC-SC (Context Sparse Constraint for Sparse Coding) algorithm is pro-posed. It employs temporal correlation of user behavior to build a candidate neighbor image set, calculates the set similarity to obtain a context optimal neighbor set, and constrains on sparse coding model with minimum distance between context neighbor images to satisfy sparseness and discrimination of sparse representation. The experimental results on commodity image classifica-tion database show that CSC-SC algorithm can improve classification accuracy effectively.