Image Classification Using Multiple Combination Features Based On Screening Sparse Coding

With the rapid development of digital technology, a variety of digital images emerge in endlessly. So how to make the computers automatically recognize the object categories using machine learning methods, has always been a very important task and challenge. In recent years, the bag-of-visual-word model plays a very important role in image classification. The method regards the local feature descriptors as the visual words, and then uses unsupervised methods such as:k-means clustering or sparse coding (SC) to train the codebook and code the features, finally uses spatial pyramid matching model for further expressing image characteristics. Despite of its worldwide popularity, there is still certain improving space in classification efficiency and computational investment when the dataset is fairly large. In this paper, we put forward two novel image classification methods according to the above problems.We first propose an image classification model using multi-directional context features and improved sparse coding algorithm (MDScSPM), which is a novel image representation method by utilizing sparse coding with screening lasso and multi-directional context features. First, a technique called Dual Polytope Projection rules (DPP) is applied during sparse coding, which is a screening method to discard the zero coefficients before the LASSO(Least Absolute Shrinkage and Selection Operator) process, and thus we can reduce the time complexity of sparse coding algorithm. Second, multi-directional context features are produced by the following steps:Firstly, we extract SIFT descriptors from sub-patches and combine each sub-patch with its neighboring regions to form super-patches context features. Then after the directional context features have been calculated, we respectively use them to do the SC using DPP screening rules and multi-scale max pooling SPM to obtain separated sparse representation of images. Specifically, to strengthen the overall stability characteristics of the image, multi-direction context features of images will be computed by the sqrt calculation function.As we can see, in the practical application, most of images contain rich color information, and in particular the samples which belongs to the same category have high color similarity. However, the traditional bag-of-visual-word model is conducted on the gray images which ignores the color information. In order to further improve the classification accuracy, we propose another new image classification method using multiple combination features based on screening sparse coding (MDScSPM+HSV). The images are converted to the HSV color space and we extract global color features of them. Finally, we combine the multi-directional context features and the global color features as the final features of the images.Experiments on several benchmarks (Caltech-101 dataset, UC Merced Land-Use dataset and UIUC-Sport dataset) show the high recognition rate and good time efficiency of the proposed methods.