Development of a Multispectral Image Database in Visible-Near-Infrared for Demosaicking and Machine Learning...

Area of Science:

• Agricultural Science
• Computer Vision
• Image Processing

Background:

• Multispectral (MS) imaging is increasingly utilized across various research domains.
• A significant challenge is the scarcity of accessible multispectral image databases due to the recent development and limited availability of MS cameras.
• The creation of comprehensive MS image databases is essential for advancing research in this field.

Purpose of the Study:

• To establish a freely accessible multispectral image database to address the limitations in current data availability.
• To provide high-resolution MS images of plants and weeds, including annotations and masks, to facilitate research.
• To support the development and evaluation of algorithms for MS image analysis, such as demosaicking, segmentation, and crop/weed discrimination.

Main Methods:

• Utilized two high-end MS cameras (visible and near-infrared) based on filter array technology from the PImRob platform at the University of Burgundy.
• Acquired and curated a dataset of high-resolution MS images.
• Provided both original raw and demosaicked images, alongside annotated images and segmentation masks.

Main Results:

• Developed and released a freely accessible multispectral image database.
• The database contains diverse MS images of plants and weeds with detailed annotations and masks.
• Included both raw and processed (demosaicked) image data.

Conclusions:

• The established MS image database serves as a valuable resource for the scientific community.
• It is particularly beneficial for research focusing on demosaicking techniques, segmentation algorithms, and deep learning applications in agriculture.
• The database aims to accelerate progress in automated crop and weed identification using multispectral imaging.