CTCNet: A CNN-Transformer Dual Branch Network for Sand Dune Image Segmentation

Dune images typically display intricate details and relatively uniform spectral characteristics, making them a unique challenge for image segmentation tasks. Due to the limitations of traditional convolution operations, the convolutional neural network (CNN)-based methods struggle to capture long-range dependencies. The Transformer-based methods have good performance in long-term dependency relationships, but it lacks modeling of local context. Based on the strengths of both approaches and the concept of a dual-branch architecture, in order to better achieve segmentation of sand dune images in desert areas, a dual-branch network combining CNN and Transformer (CTCNet) is proposed, and comparative experiments are conducted on a self-made Chinese desert sand dune morphology dataset and the Registan-Kharan desert sand dune morphology dataset. The CNN branch aims to capture local information and enhance feature extraction capabilities using convolutional attention block (CAB). The Transformer branch captures global information and employs an enhanced transformer block to improve the capture of remote dependencies and create more discriminative features. Afterward, the dual branch features are merged through a feature fusion module (FFM) to enhance the capture of finer details. Driven by its dual-branch structure and other design features, CTCNet has demonstrated exceptional capability in capturing dependencies and contextual information for sand dune image segmentation tasks. The experimental results demonstrate that CTCNet achieves an accuracy of 89.66% and a mean intersection over union (MIoU) of 82.29% on the Chinese desert sand dune morphology dataset. On the Registan-Kharan desert dune morphology dataset, it achieves an accuracy of 91.24% and a MIoU of 81.21%. Outperforming other models of similar complexity and size, achieving state-of-the-art results, and demonstrating the effectiveness and robustness of CTCNet.