Multiscale and Multitemporal Road Detection from High Resolution SAR Images Using Attention Mechanism

Road detection from images has emerged as an important way to obtain road information, thereby gaining much attention in recent years. However, most existing methods only focus on extracting information single temporal intensity images, which may cause a decrease image resolution due the use of spatial filter avoid coherent speckle noises. Some newly developed take into account multi-temporal preprocessing stage noise SAR imagery. They ignore characteristic objects such consistency for multitemporal that cover same area and are taken at adjacent times, causing limitation performance. In this paper, we propose multiscale network (MSMTHRNet) imagery, contains enhancement module (TCEM) fusion (MSFM) based mechanism. particular, TCEM make full submodule applies mechanism capture contextual information. We enforce constraint by enhanced feature representations imagery help distinguish real roads. Since width roads various, incorporating features is promising improve results detection. MSFM learned weights combine predictions different scale features. there no public dataset, build dataset evaluate our methods. State-of-the-art semantic segmentation HRNetV2 used baseline method compare with MSHRNet MSMTHRNet. The MSHRNet(TAF) whose input after adopted proposed On test MSMTHRNet over 2.1% 14.19%, respectively, IoU metric 3.25% 17.08%, APLS metric. improves MSMTHRNet(TAF) 8.23% 8.81% metric, respectively.