Arazi Segmentasyonu için Multispektral Görüntü Füzyonu

Semantic segmentation of terrain images plays a critical role in military, agricultural, and logistics applications. The fusion of images obtained from different spectral bands enables more accurate and comprehensive analyses. In this study, a dual-stream fully convolutional segmentation network is proposed, which takes both RGB and thermal images as input. The model processes each modality with independent encoders, extracts high-level features, and transforms them into a unified representation. In the transmission from the encoder to the decoder, learnable inter-convolutional connections are employed instead of traditional skip connections, ensuring a more effective fusion of RGB and thermal feature maps. As a result, significant improvements in segmentation performance have been observed, particularly under low-light and foggy conditions. Experimental results demonstrate that the proposed method achieves up to a 16.58% mIoU score improvement compared to approaches using only RGB or thermal images.