A Rapid Image Semantic Segment Method Based on Deeplab V3+

Abstract

AbstractAs one of research directions in the fields of image processing and computer vision, image semantic segmentation classifies images by pixel for semantic comprehension. Due to the lack of a general theory of segmentation, conventional image semantic segmentation algorithms are mostly on the strength of specific image models, so a variety of methods develope with their respective range for using and relative merits. What’s more, it’s a pity that most of them can only use the underlying features to generate simple segmentation results of foreground and background. Since the breakthrough in deep learning at the ImageNet competition in 2012, convolutional neural networks have taken every field of computer vision by storm. Image semantic segmentation has broken through the limits of traditional methods rapidly after the integration of deep learning. After making use of convolutional neural networks, its accuracy for segmentation has been improved effectively so that image semantic segmentation at present can further meet the requirements of segmentation technology in different application scenarios. In recent years, it has developed rapidly and gradually become a research hotspot. This paper mainly studies the image semantic segmentation methods about Deeplab series based on deep learning and raises a lightweight model for semantic segmentation based on Deeplab v3+. Regarding MobileNet as backbone to extract features, it makes use of atrous spatial pyramid pooing to gain global multi-scale features. Then the decoder structure concatenates multi-scale features and low-level features to enhance the expression for spatial features and improve its performance. After sufficient experiments on PASCAL VOC using the fast semantic segmentation model, the result reveals that it obtains a good balance between accuracy and speed.KeywordsDeep convolutional neural networksSemantic segmentationDeeplab