Two-stage Context-aware Based Monochromatic Shadow Detection

Shadow detection is to distinguish the shadow region of image,and accurately detect it.As a pre-processing task,shadow detection has been widely used in computer vision fields such as video tracking,pedestrian recognition,object detection,and semantic segmentation.Existing deep-learning based shadow detection algorithms have achieved good results in general shadow scenes and some complicated situations.However,these algorithms are usually designed for color image.In some specific task scenarios,it is necessary to perform shadow detection on monochromatic images(such as remote sensing images,infrared images,etc.).Monochromatic images usually exist problems such as low contrast differences between shadow and non-shadow regions,and difficulty in identifying black objects and their shadows compared to color images.This thesis mainly focuses on the shadow detection of monochromatic images.First,some existing shadow detection algorithms,especially three shadow detection algorithms are introduced in our work.Then,a novel algorithm of monochromatic shadow detection is designed by using the long-range dependency context.The main works of this dissertation are as follows:(1)A monochromatic shadow detection model based on two-stage context-aware is proposed to achieve accurate detection of shadows in monochromatic images.The model extracts long-range dependency context features and multi-scale spatial context features in two stages of non-local and local extraction,respectively.Different from only extracting the traditional context in a local way,long-range dependency context is extracted in a non-local way and allowed any feature points in the image to establish mutual correlation.This combination of feature extraction plays a key role in the accurate discrimination between shadow and nonshadow regions within complicated scenes.The model consists of an end-to-end encodedecode network structure.(2)In the branch of encoding network,a context extraction framework based two-stage is proposed.First,a module that extracts long-range dependency features is embedded into the pre-trained network in the non-local extraction stage.The improved backbone network can extract the features with long-range dependency context.Then,a residual dense atrous module is proposed in the local extraction stage.The module can be placed at different resolution layers to extract local spatial context features and densely fuse them.Finally,shadows within complicated monochromatic images can be accurately detected through twostage feature extraction.(3)In the branch of decoding network,a fusion module of semantic guided attention is proposed by using high-level to guide low-level.This module can suppress the noise in the context features and boost the useful hierarchical complementary information.The consistent integrity of shadow regions can be improved by using high-level features to merge and guide low-level features.(4)In addition,in order to better demonstrate the validity of the proposed monochromatic image shadow detection algorithm,a more challenging aerial shadow dataset ASD(Aerial Shadow Dataset)is constructed.The ASD dataset is obtained via collecting and filtering data,labeling ground truths,and cleaning data.It is mainly composed of 1459 training images and 209 test images,covering varieties of scenarios such as airfields,bridges,buildings,and factories.Finally,the proposed algorithm is implemented by using Pytorch under Ubuntu 16.04 environment.In addition,five state-of-the-art shadow detection methods and two salient object detection methods are compared with our proposed model.The experimental results on three graying public shadow datasets and the newly constructed ASD dataset demonstrate that our algorithm can more accurately and completely detect the shadow regions from monochromatic images,especially for those images with low contrast images and those images with shadows of black objects.At the same time,the proposed residual dense atrous module is also applied to multi-focus image fusion task and experimental results demonstrate that it has good versatility.