Gated Fusion based Multi Scale Feature Network for Landmark detection

Architectures derived from the U-Net framework and its convolutional neural network (CNN) variants have garnered significant attention in computer vision due to their remarkable achievements. However, these methods predominantly focus on leveraging high-resolution low-level features for pixel-wise tasks such as semantic segmentation. Consequently, they fail to fully exploit global features and deep semantic information, resulting in insufficient localization of target regions and diminished performance in tasks requiring both global and local feature integration. In contrast, transformer-based frameworks have demonstrated superior efficacy in high-level semantic tasks. To address these limitations, a Multi-Scale Feature Network (MSFNet) is proposed, which is a novel cross-connected architecture integrating U-Net and Transformer, with multi-scale semantic feature learning being emphasized.