Revealing the Unseen: A Single-Stage Attention Based Occluded Object Detection Model in Remote Sensing Imagery

Object detection based on deep learning has achieved promising results on traditional datasets, but detecting objects in remote sensing imagery with diverse occlusions remains challenging. This is due to the fact that natural occlusions are common in real-world images, and there is a lack of adequate datasets and neglect of latent information that can be useful for identification. Our research endeavors to tackle this challenge by presenting an innovative single-stage image-adaptive YOLO-transformer framework that leverages attention-based mechanisms to enable the model to concentrate on important regions and extract more distinctive features. To optimize the model’s accuracy while maintaining its lightweight and suitability for real-time applications, we employed a depthwise convolution and SiLU activation function in lieu of the standard convolution. These modifications allow our framework to attain high levels of precision. Our approach is evaluated on synthetically generated occluded datasets from publicly available NWPU-VHR10 and demonstrated adaptive processing of images in both normal and three types of environmental occlusions (foggy, rainy, and cloudy). The experimental results are highly promising, as our approach achieved an inference speed of 6 ms with 11.9 GFLOPs on the NVIDIA Tesla T4 while maintaining effectiveness in all three occlusion types.