Abstract:
The success of deep-learning-based algorithms significantly boosted the performance of computer vision methods based on color images in recent years.
The development of hyperspectral and multispectral camera systems allowed many new applications of spectral imaging.
Still, the combination of both, computer vision methods and spectral imaging, are in its infancy, especially since deep-learning-based approaches are not well-established for spectral imaging.
In this work, recent computer vision developments are applied to different spectral imaging tasks.
Four challenges for the algorithms (lack of data sets, task-specific features, complicated data augmentation, and large channel dimension) are identified and tackled.
In the first part of this work, a simple convolutional neural network is proposed and evaluated on two hyperspectral imaging applications in food inspection. In this context, a data set of ripening fruit is introduced, which is used throughout the rest of the work.
In the second part, self-supervised pretraining for hyperspectral imaging is introduced based on the example of three state-of-the-art contrastive learning methods (SimCLR, SimSiam, Barlow Twins). Some modifications, like data augmentations, are required for this.
Afterward, the main contribution of this paper, a wavelength-aware 2D convolution for hyperspectral imaging, is proposed. The key idea of the method is the introduced bias "Similiar wavelengths show similar features". This bias leads to a significant trainable parameter reduction and supports the training of camera-agnostic models.
The last part of this work discusses and evaluates the usefulness of multispectral cameras for maritime search and rescue missions. Therefore, a data set with humans in open water was recorded and published. In this context, a method is presented which can reduce the background bias, a problem of these remote sensing recordings.
In the end, the work is concluded with a summary, a short discussion, and an outlook.
None of the defined challenges were fully overcome. Still, the presented approaches show how a solution could look and prepare future research in these directions.