Loading...

Attention-based Change Detection and Panoptic Change Segmentation

Ebrahimzadeh, Mohammad | 2024

170 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57380 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Manzuri Shalmani, Mohammad Taghi
  7. Abstract:
  8. Change detection involves comparing images of a specific area taken at different times to identify the key changes that have occurred over time. It is critically important in the field of remote sensing, with applications spanning urban planning, disaster monitoring, and environmental conservation. However, challenges such as varying weather conditions and lighting in satellite images taken at different times, as well as the presence of non-target changes, can make this task challenging. In this research, we will focus on changes in land-use, particularly in building infrastructure, which is crucial for urban development planning. Moreover, since change detection does not specify the type of change (e.g., construction or demolition of a building) or the number of changed instances, we propose panoptic change segmentation by combining this task with the task of panoptic segmentation. This approach can provide more comprehensive information than conventional change detection. With the advent of transformers from the field of natural language processing to computer vision, we have witnessed promising advancements in many existing tasks in this domain. These advancements demonstrate the capability of the transformer model and its self-attention mechanism as the core building block of its architecture. However, all previous transformer-based methods for the change detection task use the common approach in semantic segmentation, which involves separate classification of each pixel. In this research, inspired by the mask classification approach—an innovative method in image segmentation—we introduce the first end-to-end deep neural network capable of performing change detection in bi-temporal remote sensing images. This network predicts a set of binary masks along with the “change” or “no change” class label. Notably, our model achieves state-of-the-art performance on the LEVIR-CD and S2Looking datasets, with F1 scores of 92.17 and 68.4, respectively, outperforming many previous change detection methods. Additionally, we present another model for panoptic change segmentation in satellite images. This model not only detects changes but also efficiently predicts the type of change and its various instances
  9. Keywords:
  10. Deep Neural Networks ; Change Detection ; Remote Sensing ; Mask Classification ; Image Segmentation ; Dataset

 Digital Object List