Research On Shadow Detection And Removal In Remote Sensing Images

With the rapid development of earth observation technology, satellite remote sensing has come into an unprecedented new stage. The high-resolution remote sensing images make people exploring and understanding the nature into a new milestone. However, there are large size shadows in the high-resolution remote sensing images since the sun’s rays are blocked by the huge buildings and trees. The existence of shadow will inevitably affect interpretation of remote sensing image and bring many difficulties to the follow-up of remote sensing image processing, such as target classification and recognition, image registration and other tasks. To effectively use of the shading data, remote sensing image processing shadow becomes a research hotspot. The shadow detection and shadow removal are two aspects of the shadow processing interdependent.This thesis studies the existing shadow detection algorithm proposed by A Makarau et al. titled"Blackbody Radiation Model" and proposed by Liu J et al. titled "Adaptive Feature Selection". The study found that the existing shadow detection algorithm has good detection performance for the luminance uniformity and low brightness of shaded area. However, due to the terrain complexity, there are many non-homogeneous shadows and light shadows in remote sensing images. This shadow often leads to omitted errors using current shadow detection methods. To improve detection performance of the intensity inhomogeneous shadows and bright shadows in the remote sensing images, this thesis proposes a shadow detection method combined level set with the feature of color space. Based on the non-uniformity of brightness of shadow regions in remote sensing images, the local classification level set segmentation model is firstly adopted to obtain the preliminary shadow regions. Then the color feature difference between shadow regions and green ones is analyzed and used to distinguish green space from the preliminary shadow areas. Then the green regions detected mistakenly are eliminated. Experimental results show that the proposed method is superior to the existing methods such as blackbody radiation model and adaptive feature selection. The omitted error of the proposed method is low since the inhomogeneous shadows and bright ones are effectively detected, and the whole detect process of it is done without manual intervention.Based on the shadow can be pinpointed by the shadow detection algorithm, this thesis researches on the shadow removal algorithm. The existing shadow removal algorithm, such as "Color-invariant"、"Example-Based learning" and "Adaptive Nonlocal Regularized", are through learning finds that these methods fail in removing shadows that are cast on curved surfaces, as well as retaining the original texture of the image, or appearing other questions like loss of feature information, false color tone, and too much manual work in operating. This thesis proposes a novel shadow removal approach based on the curvelet wavelet transform. This algorithm designs the three strategies to improve the performance of the "Adaptive Nonlocal Regularized".(1) Shadow detection method combined level set with the feature of color space to extract the shadow area;(2) Shadow edges weakened to reduce the influence of the penumbra;(3) Shadow area direction factor is extracted by curvelet wavelet for texture-direction nonlocal regularized. The experimental results show that this method improves the operability algorithm and get better textural information than the shadow removal scheme based on adaptive nonlocal regularized.This thesis develops a GUI demonstration system based on MATLAB platform and provides the instruction and demonstration results of the system as well.