QSU-Net: QCNN with Self-attention based U-Net for Single Image Dehazing
REVANTH BANALA, Manoj Kumar, Sanjay Kumar Dwivedi · Research Square · 2024
Abstract Single-image dehazing refers to the procedure of eliminating haze elements from a provided image that contains haze. Dehazing is a complex and ill-posed problem in the field of computer vision. This article introduces a new QSU-Net, which is a hybrid layered U-Net with attention-based skip connections. The purpose of this network is to remove haze from single images. The QSU-Net's hybrid layer, also known as the network layer, comprises a sequence of Quaternion Convolutional Neural Networks (QCNNs) and attention blocks. A QCNN block of three layers. The first layer is the QCNN layer, followed by the Rectified Linear Unit (ReLU) layer, and lastly another QCNN layer. The attention block consists of a self-attention block, which includes convolutional neural network (CNN) layers at the lower part and multi-layer perceptron (MLP) layers at the upper part to handle attention. The U-Net model is trained using three widely recognized datasets: I-HAZE, O-HAZE, and Dense-HAZE. The training process involves utilizing a L1 loss function for a total of 500 epochs. The efficacy of the proposed U-Net is subsequently assessed by contrasting its outcomes with alternative techniques for image dehazing. The evaluation metrics used include peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), natural image quality evaluator (NIQE), realness index (RI), and visibility index (VI). Our solution consistently surpassed other dehazing methods in terms of performance.