Research And Implementation Of Semantic Segmentation Method Based On Image Enhancement

In recent years,with the increase in data volume and computing power,especially the large-scale use of GPU computing,Deep Learning has gradually established its dominant position in the field of computer vision.In many sub-fields,including image classification,image segmentation,image detection and speech recognition,deep neural networks(DNNs)have achieved the best results.Among many deep neural network structures,convolutional neural network(CNN)is the most famous one.Almost all computer vision problems use this structure,which can be trained using large-scale data.CNNs use operations like shared weights,pooling,and random dropout to reduce computation and improve the generalization capabilities.Semantic segmentation is one of the most fundamental tasks in computer vision.It has important applications in the fields of driverless car,medical imaging,geographic remote sensing,robot navigation,etc.Its mission goal is to make every pixel of input image classified.In recent years,deep learning has showed outstanding performance on such intensive labeling problems.However,recent the semantic segmentation methods based on CNNs focus on how to better fuse the features of a single input image.Few researchers pay attention to how to segment an image more precisely by enriching the features of the image itself.This thesis will introduce and discuss several classical CNN-based semantic segmentation frameworks,summarize their structures,work principle and characteristics,and construct enhancement network and attention network to boost existing segmentation frameworks from the perspective of image enhancement.We can achieve better segmentation performance by proposed method.This paper mainly includes the following three parts:1.This paper proposes an image enhancement network that can simulate image enhancement algorithms and help semantic segmentation tasks.2.This paper proposes an attention network that can produce more conducive images to semantic segmentation task.3.This paper builds an end-to-end pipline by combining the enhancement networks,attention network and existing segmentation network and improve the semantic segmentation performance of baseline model.