A Study Of Mid-level Semantic Feature Learning In Image Recog- Nition

The rapid development of mobile electronic terminals and 4G Internet have given people a great convenience to use the Internet apps in their daily life. At the same time, the data processing ability of digital information collection devices is further strengthened, the common social tools have been able to support large and complex image and video processing, which makes people more depend on the image and video data in their daily and social life. Data analysis based on text processing has not been able to effectively reflect the data content, image and video data analysis has become an urgent problem to be solved.The most difficult problem in image and video analysis lies in the understanding of its senmantic content. Image understanding can be divided into three levels, which contains visual layer, object layer and conceptual layer. Visual layer features cannont effectively describe the semantic content of images and conceptual layer information is the goal of image understanding that the researchers are working on. This paper focuses on the object layer feature learning and two works have been done to learn effecient mid-level semantic features.Firstly, this paper proposed a multi-level feature representation for visual concept detection. This work focuses on the latent semantic relation between low-level feature descriptors, which is learned by topic model. This infomation is used as the mid-level semantic feature model and we also take the low-level infomation into account. Fisher vector is used to learn the global feature representaion based on the low-level features. In the experiment, we fuse this two different feature representaions and we validate the effectiveness of the proposed feature representation through the semantic concept detection task.Secondly, this paper also proposed a mid-level semantic feature learning method based on the optimal frequent patterning mining. This work paied special attention to the co-occurence of low-level feature descriptors and use it as mid-level semantic feature model, specifically, the optimal pattern model is taken into use. We first mine useful patches from the input image dataset, and then mine pattern model from this weak semantic patches. At last, we learn the image's representation based on this mid-level semantic feature model and test the validity through a image classification task.