Research On Real-time Video Segmentation Technology Based On Deep Learning

Video segmentation is a hot topic in the field of image processing and computer vision.It is a technology to separate the regions of interest from video and has been widely used in identity recognition and intelligent monitoring.In the applications of object detection and object segmentation,deep learning has great advantages over traditional algorithms in both accuracy and speed.Therefore,it is of great significance to research the video segmentation theories based on deep learning.At present,with the increasing demand for photography on smartphones,tablets and other portable devices,array cameras have better development prospects.In this paper,3 × 3 microarray camera was used to obtain images,which has wider field of vision and more image details compared with the traditional cameras.Besides,microarray camera has shorter shutter time intervals,so it can obtain images with better time continuity.Based on the current deep learning algorithm,experimental research was carried out,and video segmentation was implemented by combining object detection and object segmentation.Compared with other object detection algorithms,Single Shot MultiBox Detector(SSD)algorithm has advantages in both speed and accuracy,but has higher complexity.Therefore,we are focused on optimizing SSD algorithm to further improve the accuracy and speed.The main work of this paper is as follows:1.Person images based on the improved YCrCb color space were added to the Pascal VOC dataset,which enhanced the feature extraction effect of person images under high illumination.2.3 x 3 microarray camera was used to capture images,and zhang calibration algorithm was applied to calibrate 9 array lenses to obtain the parameters of each lens.The results of video segmentation are more continuous than traditional cameras.3.We proposed a video segmentation algorithm based on SSD algorithm,which was compared with the existing deep learning network.The prediction process of SSD algorithm was improved to reduce the number of prediction boxes and the number of iterations in the Non-Maximum Suppression algorithm.Furthermore,the VGG network model was improved,in which the superposition of small-scale convolution kernel is adopted to replace 3 x 3 convolution kernel,reducing the number of parameters in the convolution layer and improving the detection efficiency.The proposed algorithm has been effectively verified in terms of segmentation speed and accuracy.The accuracy of our model is higher than that of SSD algorithm by 2.3%,and the segmentation efficiency is improved as well.