Two-stage Domain Adaptation for Infrared Ship Target Segmentation

Ship target segmentation in infrared scenes has always been a hot topic, since it is an important basis and prerequisite for infrared-guided weapons to reliably capture and recognize ship targets in the sea-level background. However, given the small target and fuzzy boundary characteristics of infrared ship images, obtaining accurate pixel-level labels for them is hardly achievable, which brings difficulty to train segmentation networks. To improve the segmentation accuracy of infrared ship images, we propose a two-stage domain adaptation method for infrared ship target segmentation (T-DANet), where the segmentation model is trained using visible ship images with clear target boundaries. In this case, the source domain is the labeled visible ship images, while the target domain is the unlabeled infrared ship images. Specifically, in the first stage, we use an image style transfer network to convert the infrared ship images into those with visible light style, so that the visual disparity between the two domain images can be reduced. Next, the visible, infrared, and converted infrared images are input into the Deeplab-v2 segmentation network for training, thereby obtaining the initial network weights. At this time, random attention modules are added separately to the low- and high-level spaces of Deeplab-v2, in order to improve its feature extraction capability. In the second stage, we mix the visible and infrared images through region mixing to acquire the mixed domain images, as well as their corresponding labels. Subsequently, Deeplab-v2 is further trained using the mixed domain images to attain better segmentation accuracy. Experimental results on both the home-made visible-infrared ship (VI-Ship) image dataset and the public infrared image dataset are superior to those existing mainstream methods, demonstrating its effectiveness