Hybrid Vision Transformers and CNNs for Enhanced Transmission Line Segmentation in Aerial Images

This paper presents a novel architecture for the segmentation of transmission lines in aerial images, utilizing a hybrid model that combines the strengths of Vision Transformers (ViTs) and Convolutional Neural Networks (CNNs). The proposed method first employs a Swin Transformer backbone (Swin-B) that processes the input image through a hierarchical structure, effectively capturing multi-scale contextual information. Following this, an upsampling strategy is employed, wherein the features extracted by the transformer are refined through convolutional layers, ensuring that the resolution is maintained, and spatial details are recovered. To integrate multi-level feature maps, a feature fusion module with a squeeze-and-excitation (SE) layer is introduced, which consolidates the benefits of both high-level and low-level feature extractions. The SE layer plays a pivotal role in augmenting the feature channels, focusing the model's attention on the most informative features for transmission line detection. By leveraging the global receptive field of ViTs for comprehensive context and the local precision of CNNs for fine-grained detail, our method aims to set a new benchmark for transmission line segmentation in aerial imagery. The effectiveness of our approach is demonstrated through extensive experiments and comparisons with existing state-of-theart methods.