Image Semantic Segmentation Of Depth Feature Fusion

In twenty-first Century,in the age of information technology,computer performance has made great progress.In the era of big data,every field has begun to attach importance to deep learning.Computer vision is widely applied to everyday life,such as driverless,face recognition,pedestrian detection and tracking,medical image detection,video surveillance,event detection and so on.Image semantic segmentation is a basic task of computer vision.If we can quickly determine the category of each pixel in the image at low computational cost and time cost,many computer vision tasks will become very resolved.In this thesis,we introduce the development process of deep learning,the convolution neural network(CNN)algorithm principle,and the structure and characteristics of the Fully Convolutional Network.This paper makes a thorough research and comprehensive experiment from two key technologies to improve the performance of semantic segmentation.Both the innovation of the deep learning algorithm and the improvement of the structure of the convolutional neural network are also presented.Finally,the performance of semantic segmentation is greatly improved.The main innovation points of this paper are as follows:we propose a fusion coefficient learning method that can guide us to select effective layers.What's more,our approaches can be added to other works that require multi-scale fusion to further boost their performance.We proposed three principles for preliminary screening of layers and presented the fusion coefficient learning algorithm.Besides,we defined a method of measuring effective receptive field.We designed a Dense Global Context Module,which makes the effective receptive field coverage larger and density higher.With the Dense Global Context Module,segmentation model reduces a large number of parameters while the performance has been substantially improved.