Visual Feature Extraction Based on Region of Interest and FCNN
Xingjie Wang, Mingtan Sun, Jiachen Pan · Advances in transdisciplinary engineering · 2023
At present, image and video has become more and more the main form of multimedia expression, how to effectively locate the user really need from the large-scale image video data image block, has become a more popular problem in the field of image and video processing, extraction of interest area is the key technology to solve this problem. Based on this, many visual attention models have been proposed, the most representative of which is the extraction algorithm proposed by and. The algorithm first extracts the features of the image such as brightness, color and direction into the feature map, and then uses the “center surround” mechanism similar to the field of perception for each feature map. In recent years, with the rapid development of three-dimensional display technology, more and more elements are integrated into the daily life. Compared with the previous flat display, three-dimensional display technology can often bring the viewer a higher quality of visual experience and lifelike visual feelings, so it is very popular. Due to the addition of depth information, the traditional image-based region of interest extraction method can not predict the salient region in stereo video very well. In this paper, we study the mechanism of human visual attention, adopt the bottom-up method, combine the traditional and video sequence motion features, comprehensively consider the depth information of stereoscopic video, introduce the perception characteristics of depth of human visual system into the region of interest extraction, and propose a visual attention model of stereoscopic video. Experiments show the effectiveness of our model.