基于MADSUNet网络的图像阴影检测OA

In shadow detection within images under complex lighting conditions,the Unet++image segmentation model may lose shallow-level details(such as texture and brightness gradients)due to stacked convolutions and downsampling operations,thereby affecting the precise localization of shadow boundaries and the detection performance in weak shadow regions.To address this issue,we propose a shadow detection model named MADSUNet.An adaptive multi-head masked attention module is incorporated into the deep encoder,which enhances the focus on key features of shadow regions through a dynamic weight allocation mechanism while suppressing interference from non-shadow backgrounds,thereby improving the model's robustness under complex lighting conditions.Additionally,an efficient dynamic upsampler is introduced in the decoder to further refine the smoothness and continuity of shadow boundaries.Experi-mental results demonstrate that the proposed model achieves high accuracy and detection performance in shadow detection.On the SBU,UCF,and ISTD datasets,the Balanced Error Rate(BER)values reach 4.99％,8.72％,and 2.06％,respectively,indicating that the model can accurately identify shadow and non-shadow regions under varying lighting conditions,effectively distinguishing shadows from dark backgrounds.