Guided U-Net Aided Efficient Image Data Storing with Shape Preservation

Abstract

The proliferation of high-content microscopes ($\sim$32 GB for a single image) and the increasing amount of image data generated daily have created a pressing need for compact storage solutions. Not only is the storage of such massive image data cumbersome, but it also requires a significant amount of storage and data bandwidth for transmission. To address this issue, we present a novel deep learning technique called Guided U-Net (GU-Net) that compresses images by training a U-Net architecture with a loss function that incorporates shape, budget, and skeleton losses. The trained model learns to selects key points in the image that need to be stored, rather than the entire image. Compact image representation is different from image compression because the former focuses on assigning importance to each pixel in an image and selecting the most important ones for storage whereas the latter encodes information of the entire image for more efficient storage. Experimental results on four datasets (CMATER, UiTMito, MNIST, and HeLA) show that GU-Net selects only a small percentage of pixels as key points (3%, 3%, 5%, and 22% on average, respectively), significantly reducing storage requirements while preserving essential image features. Thus, this approach offers a more efficient method of storing image data, with potential applications in a range of fields where large-scale imaging is a vital component of research and development.