Implementing W-Net deep learning for terahertz image enhancement and segmentation

Abstract Terahertz images typically suffer from poor image quality and do not allow traditional machine learning methods to detect and identify objects. In this work, an image segmentation algorithm called W-Net is deployed for the first time, to segment concealed objects from the poor-quality Terahertz image samples. This involves developing an image enhancement and segmentation system for terahertz images. The proposed system comprises of two stages: in the first stage, the resolution of low resolution THz images has been enhanced by training a U-Net deep learning model. In the latter stage, another U-Net model is trained to segment the enhanced images using corresponding masks to identify regions of interest. The proposed system is evaluated using a dataset of low-resolution terahertz images with its respective high resolution images and masks. The system performed well with an accuracy, precision, receiver operating characteristics and F1 scores of 99.1%, 0.9143, 0.9977 and 0.9146, respectively for enhancement. Moreover, values of the same evaluation metrics deduced as 99.9%, 0.9851, 0.9999 and 0.9923, respectively for segmentation. Therefore, the demonstrated scheme is expected to identify and classify objects hidden in real-time terahertz images and improve the quality of the same to enable accurate segmentation for applications like security screening, biomedical applications, quality and health monitoring in food industry, etc.