Shadow Detection Based On Deep Learning

As one of the important research directions in the field of computer vision,shadow detection aims to accurately identify shadow areas in images,which plays a crucial role in high-level image processing applications such as inferring light fields,video tracking,object detection,and salient target detection.In recent years,with the rapid development of deep learning techniques,shadow detection methods based on deep learning have made breakthrough progress.However,these methods still have some limitations,such as complex network model structures,a huge amount of parameters,and excessive reliance on large-scale training data and corresponding pixel-level shadow labels.To address such problems within the existing shadow detection methods,this thesis proposes two shadow detection algorithms and conducts several experiments on some public shadow detection datasets to verify their effectiveness.The details of this thesis are as follows:(1)Considering that existing deep shadow detection algorithms usually have complex network models and a huge amount of parameters,and are easy to overfit when the training data sizes are too small,this thesis proposes an image shadow detection algorithm based on a lightweight network.The proposed algorithm uses a small-scale network as the backbone network for feature extraction and replaces the standard convolution with the deep separable convolution,which has lower computational complexity than the traditional convolution.In addition,to compensate for the performance reduction caused by lightweight,a spatial perception module and a semantic perception module are specially designed to mine the spatial detail information and semantic context information for shadow detection.As well,a feature-guided fusion module is designed,which uses the complementarity of features at different levels of the network model,to achieve feature fusion for further improving the detection performance of the model.(2)To address the problems that existing deep shadow detection methods rely excessively on large-scale training data with pixel-level labels,this thesis explores unsupervised learning for shadow detection and proposes a shadow detection method based on deep unsupervised learning.The method combines self-training and curriculum learning for predicting shadows from the pseudo-labels generated by some traditional shadow detection models.The deep model is guided through curriculum learning to progressively learn shadow knowledge from simple samples to complex ones,and a pseudo-label update mechanism based on the selftraining method is used to progressively update the pseudo-label during the curriculum learning process to reduce the noise and improve the robustness of the model against noisy pseudo-labels.The two proposed algorithms are implemented by using Pytorch under Ubuntu 18.04 environment.For the Lightweight Network shadow detection algorithm,the experiments are analyzed and compared with existing deep shadow detection algorithms on some publicly available shadow detection datasets.The experimental results demonstrate the superiority and effectiveness of the proposed algorithms.The proposed deep unsupervised learning for shadow detection algorithm is compared with existing shadow detection algorithms on some publicly available shadow detection datasets.The experimental results demonstrate that the proposed algorithm is superior to the existing conventional unsupervised shadow detection methods and even achieves comparable performance with some shadow detection models based on supervised learning.