Research On Image Content Recognition Based On Sparse Coding And Machine Learning

With the development of the technology of digital image and Internet, there are huge amounts of image generated and shared. It became more and more important to organize image data effectively, which will enable users to search their desired images efficiently. In this context, image content recognition has become one of the hottest research topics in computer vision. Image recognition is a process which calculates, analyzes and learns these input images and outputs the result of objects recognition, the relationship among the objects, scene recognition and so forth. In short, image recognition is realized by computer for understanding image in a way as human visual. In addition, the application of image recognition is not limited to content-based image retrieval. There are a lot of other applications of image recognition in many fields such as robot vision, remote sensing image recognition, medical image recognition, biometric identification and so on.This paper firstly presented conclusive analysis of related research work on image recognition globally in recent years. Then this paper emphasized on image content classification, object recognition and localization based on sparse coding model and machine learning.A novel image classification algorithm based on spatial sparse coding model and random forests is proposed in this paper. Firstly SIFT descriptors are extracted from images. Then sparse coding theory is adopted to train a dictionary and convert SIFT descriptors into sparse vectors. An efficient pooling method is employed to merge the sparse vectors in each grid of image and a pooled sparse vector is formed to represent the grid. According to the position of grids, the pooled sparse vectors are combined to form a sparse vector for representing an image. Secondly random forests multiclass classifier is employed to classify sparse vectors of images. The experiments were conducted on Caltech-101 and Scene-15 dataset and experimental results show that this algorithm outperforms several state-of-the-art algorithms.A novel image content classification, object recognition and localization algorithms based on image segmentation, sparse coding model and multiple-instance learning is proposed in this paper. The algorithm employed the concept of multiple-instance learning to treat images as bags, sparse features which are converted from SIFT descriptors as instances and visual vocabulary generated by using sparse coding model as feature space. Then the amounts of instances in bags are mapped to the feature space. Next, 1-norm SVM is employed to classify images and distribute weights to select important instances corresponding to sparse features. Finally instances are classified to localize objects using image segmentation. The experiments were conducted on Caltech-101 and Scene-15 dataset and experimental results show that this algorithm performs well both in image classification and object localization.