Font Size: a A A

Research On Low-Light Image Algorithm Based On Attention Mechanism

Posted on:2021-04-28Degree:MasterType:Thesis
Country:ChinaCandidate:L WangFull Text:PDF
GTID:2518306308474264Subject:Information and Communication Engineering
Abstract/Summary:PDF Full Text Request
Low-light image enhancement technology is an essential challenge in low-level computer visual tasks,which aims to prompt the low-light images into resolvable and clear ones.Output images can be used for subsequent higher-level application,such as object recognition,object detection,semantic segmentation,etc.In recent years,with development of deep learning and improvement of computer performance,low-light image enhancement algorithms based on deep learning have received widespread attention.However,various factors,such as environmental brightness,camera shake and object movement,pose great challenge to the topic.This paper first introduces the background knowledge and main research contents of low-light image enhancement.Through in-depth study of existing algorithms,a novel algorithm based on the attention mechanism and multi-level feature fusion is proposed to tackle image enhancement task.In a nutshell,main contributions are as follows.First,this paper proposes a channel-wise attention module,which aggregates the global information of network features using global average pooling.The module obtains the dependency relationship between the feature channels based on squeeze-and-excitation networks,assigning weights to the feature channels.The attention mechanism enhances useful information and suppresses useless information.Secondly,this paper proposes a feature extraction network and a top-down feature fusion network.The feature extraction network is used to extract multi-level feature vectors.The top-down feature fusion network over channel dimension is used to fuse different levels of features.Learning the information of feature maps at different levels enhances feature generalization and expression ability of the algorithm.In this paper,extensive experiments are conducted on three low-light image datasets,the synthetic,S7ISP,and SID datasets.The experimental result shows that the proposed algorithm outperforms the state-of-the-art algorithms.At the same time,the ablation experiments are performed on the proposed algorithm,and the result proves that the improvement of each module in the algorithm contributes to the overall algorithm performance.
Keywords/Search Tags:attention mechanism, low-light image enhancement, multi-level feature fusion, convolutional neural network
PDF Full Text Request
Related items