Research On Local Semantic Concept Representation Based Image Scene Classification Technology

In recent years, with the rapid development of computer technology, communication technology and the Internet technology, multimedia data grows explosively and a massive information environment is generated. Facing the massive image data, how to make the computers automatically understand those images and classify them into different semantic categories according to human cognition, then classify and manage the huge amounts of image resources quickly and effectively becomes an important problem urgently needed to be solved in the field of image research. Image scene classification automatically annotates images with a group of semantic labels according to the semantic content included in the whole images or image regions. Image scene classification can support semantic based image analysis and retrieval, meanwhile helps to provide effective context for image understanding on higher layer. The key problem of image scene classification is how to eliminate the semantic gap between low-level visual features and high-level semantic concept. By extracting the image local invariant features, employ the local semantic concept representation is an important research idea. This paper mainly researches on local semantic concept representation based image scene classification technology. Based on extracting the local features of images, this paper respectively researches on visual vocabulary model based, sparse coding model based and semantic topic model based image scene classification technology. The novelty and main contributions are listed in the following five aspects:1. To overcome the synonymy and polysemy problem of visual words, this paper proposes a LSI and soft-weighting based image scene classification algorithm. Firstly, latent semantic indexing technology is employed to mine the latent semantic relationship of visual words, which conducts dimensionality reduction on the large-scale visual vocabulary to obtain the compact semantic visual vocabulary. Secondly, a soft-weighting scheme is implemented to realize the mapping of local features to visual words, which maps feature points to multiple neighbor visual words with different weights. Then, make statistics on visual words’ appearance frequency in the image and construct the visual vocabulary distribution histogram representation. Finally, SVM classifier is utilized to perform scene classification. Experimental results demonstrate that the novel algorithm can effectively solve the problem of visual word synonymy and polysemy and improve scene classification performance.2. Aim to the intra-class diversity problem of images, this paper proposes an E2LSH-MKL based image scene classification algorithm. Firstly, E2LSH algorithm is employed to perform clustering to construct visual dictionaries and produce E2LSH based visual vocabulary distribution histogram representation. Secondly, by combining the advantages of non-linear multiple kernel combination methods, a nonlinear and non-stationary multiple kernel learning method—E2LSH-MKL is constructed. E2LSH-MKL utilizes Hadamard product to realize nonlinear combination of multiple different kernels in order to make full use of information generated from the nonlinear interaction of different kernels. Meanwhile, E2LSH-MKL employs E2LSH based clustering algorithm to group images into sub clusters, then assigns cluster-related weighting of multiple kernels weights according to relative contributions of different kernels on each image subset thereby realizing non-stationary weighting of multiple kernels to improve learning performance. Finally, the E2LSH based image visual vocabulary distribution histogram representation and E2LSH-MKL classifier are combined to perform image scene classification. Experimental results demonstrate that E2LSH-MKL based image scene classification algorithm performs superior to other related multiple kernel learning algorithms and is effective in resolving the intra-diversity problem of images.3. To Overcome the drawbacks of spatial information lack and weak discrimination, this paper propose an image scene classification algorithm based on sparse coding with fisher discriminative criterion constraint. Firstly, the non-negative sparse locally linear coding is constructed to encode the local features with their neighbor visual vocabularies, thus to make full use of images’spatial information. Secondly, on the basis of the non-negative sparse locally linear coding, fisher discriminative criterion constraint is added to construct a non-negative sparse locally linear coding model with fisher discriminative criterion constraint, thus to obtain the discriminative sparse representation of images. The novel model can promote the spatial separability of sparse coefficients and enforce the classification capability of images’sparse representation. Finally, SVM classifier is combined to perform scene classification. Experimental results show that our algorithm efficiently utilizes spatial information of images and incline to seek images’ discrimination representations, thus performs superior to other related algorithms and is more suitable for image classification tasks.4. An appropriate number of topics in pLSA model is important but difficult to select, in order to adaptively select the best number of topics and inspired by the theorem that the model reaches optimum as the average similarity among topics reaches minimum. This paper proposes a method of adapatively selecting the best pLSA model based on density. Experimental results demonstrate that this algorithm can achieve the performance matching the best of pLSA without manually tuning the number of topics.5. To efficiently utilize image multi-scale and contextual semantic information, this paper proposes a novel image scene classification algorithm based on multi-scale and contextual semantic information. Firstly, Images are decomposed into variant scales and diverse visual details are extracted from different scale layers. Secondly, a density-based adaptive selection method is employed to choose the best topics number for probabilistic latent semantic analysis model. Then, the pLSA model and Markov random field are combined to mine the contextual semantic co-occurrence information of image patches, thus to construct more accurate visual words. Finally, make statistics on the frequency of visual words in diverse scale layer and linearly combine them to form a multi-scale histogram as the image representation which is subsequently used in scene classification with SVM classifier. The experimental results demonstrate that our novel algorithm effectively utilizes the multi-scale and contextual semantic information of images and improves image scene classification performance.