Research And Application Image Semantic Segmentation Based On Deep Learning

Image Semantic Segmentation is a very important task in the field of Computer Vi-sion,which aims to classify every pixels to its corresponding object class.As it provides both high-level classification and low-level pixel position information,image semantic segmentation plays a crucial role in image analysis and also support the complex applica-tions,like background separation,human parsing,scene understanding and autonomous driving.The critical point in image semantic segmentation is giving the high-level classifi-cation to every pixels.Objects from one class may present different color,texture and shape;while different type of objects may contain similar local patches.Traditionally,people design hand-crafted features to differentiate these conditions,which is suboptimal.With the rise of deep learning,hand-crafted features are gradually replaced by data-driven features,and the deep learning itself becomes the standard toolkit in computer vision,too.Combining the deep learning and image semantic segmentation,this paper presents the following works:1.Based on the "Fully Convolutional Network",we build the baseline model for succeeding experiments.We improve the fundamental network and introduce the lower-level feature map in our baseline model.The performance is better than original ones.2.Combining the characters of image semantic segmentation and the method of image classification,we propose the "Global Convolutional Network".It solves the lack of "valid receptive field" in deep learning,and guarantees least parameters with best performance among other candidates.3.According to the analysis on the accuracy near object boundaries,we propose the"Boundary Refinement Network",which tries to revise the blurred object boundaries by residual learning and results int better visual effects.4.We implements the above structure design in different datasets and gets the substan-tial improvements,which verifies their validity and generality.