Tuesday 11 March 2025
The transfer of labels between multispectral cameras has long been a challenge for researchers and engineers working in fields such as agriculture, medicine, and remote sensing. These specialized cameras capture images across multiple frequency ranges, providing valuable insights into the world around us. However, the process of manually labeling each image can be time-consuming and labor-intensive.
Recently, a team of researchers from the University Institute for Computer Research has developed a new method to automatically extend bounding box (BB) and mask labels across different channels on multispectral cameras. By combining phase correlation with a refinement process, their approach enables the efficient transfer of labels between cameras, allowing for more accurate and efficient analysis of images.
The team’s method begins by using phase correlation to align images from different channels. This involves smoothing the images, transforming them into the frequency domain, and then applying cross-power spectrum formulas to retain only the phase information. The resulting peak location is used to determine the translation between the two analyzed images.
To refine this process, the researchers developed an iterative algorithm that searches for a better transformation within a proximity window. This ensures that the best possible transformation is saved, providing the highest percentage of intersection over union (IOU) metric.
The team tested their method using a MicaSense RedEdge-MX Dual camera, which captures images across 10 multispectral bands. By labeling just 12 images from band 5 with high contrast, they were able to obtain transformations for BB and mask label types with an accuracy of over 97% and 94%, respectively.
To demonstrate the effectiveness of their approach, the researchers generated artificial RGB images using the inverse of the obtained transformations. These fake RGB images allowed them to perform labeling in colored images, which can be particularly useful in applications such as agricultural monitoring or medical imaging.
Future work will focus on testing this method with more multispectral cameras having different morphologies, as well as exploring the potential benefits of incorporating a RGB camera to avoid generating fake RGB images and accumulating small errors. Additionally, the team plans to create a dataset of domestic waste for training deep neural networks and evaluating their performance using 12-bit images.
This research has significant implications for various industries that rely on multispectral cameras, such as agriculture, medicine, and remote sensing. By automating the labeling process, researchers can focus more on analyzing the data and less on tedious manual labeling tasks. The potential benefits of this approach include improved accuracy, increased efficiency, and reduced costs.
Cite this article: “Automated Label Transfer between Multispectral Cameras using Phase Correlation”, The Science Archive, 2025.
Multispectral Cameras, Automatic Labeling, Image Alignment, Phase Correlation, Refinement Process, Bounding Boxes, Mask Labels, Remote Sensing, Agriculture, Medicine
