Semantic Segmentation Of PolSAR Images Based On Complex-valued U-Net

Synthetic Aperture Radar(SAR)is an earth observation system that can quickly acquire target information and image the target with high resolution,and has the ability of all-weather and all-day operation,which has been widely used in military and civil fields.As one of the research hotspots in the field of SAR,SAR image interpretation technology can accomplish the tasks of detection,identification and classification of targets in SAR images.In recent years,with the rapid development of deep learning,the image semantic segmentation technology in the field of optical images has made remarkable achievements,and this technology has also promoted the development of SAR image semantic segmentation technology.Since the SAR imaging mechanism is different from that of optical images,SAR images are affected by coherent speckle noise.In addition,SAR image data acquisition is more difficult compared with optical images.Therefore,some optical image semantic segmentation techniques based on deep learning are difficult to achieve the ideal segmentation performance on SAR images.In this paper,we combine the characteristics of SAR images,extend the real image semantic segmentation network U-Net to the complex domain,and improve the complex U-Net by using Generative Adversarial Network(GAN)and complex capsule network,respectively,to finally realize the semantic segmentation of Pol SAR images.The details are as follows.1)A network model of image semantic segmentation based on complex U-Net and GAN is proposed.The main framework of this network model is GAN,where the generator consists of complex U-Net and the discriminator consists of real multi-resolution convolutional neural network.During the network training process,the complex U-Net achieves the initial image semantic segmentation,and the GAN further makes the image semantic segmentation results close to the real label values.The experimental results of two Pol SAR datasets show that the methods based on complex U-Net and GAN can obtain higher segmentation performance than complex U-Net.2)An image semantic segmentation network model based on complex U-Net and complex capsule network is proposed,i.e.,a complex capsule network is introduced between the encoding part and the decoding part of the complex U-net network.Since the commonly used capsule networks are real networks,this paper first proposes a complex dynamic routing mechanism to extend the real capsule network to the complex domain.Then,the complex U-Net and the complex capsule network are combined to achieve the semantic segmentation of Pol SAR images with few samples.The experimental results of two Pol SAR datasets show that the methods based on the complex U-Net and the complex capsule network can obtain higher segmentation performance than the complex U-Net.