MTS-CSNet: Multiscale Tensor Factorization for Deep Compressive Sensing on RGB Images

A team of researchers introduced the MTS-CSNet, a novel deep learning framework for compressive sensing of RGB images, in a technical paper submitted to the arXiv preprint repository on February 12, 2024, to address the scalar and receptive field limitations found in traditional convolutional neural networks. The proposed system utilizes Multiscale Tensor Summation (MTS) factorization to provide a more structured approach to multidimensional signal processing. This advancement aims to improve how high-resolution visual data is sampled and reconstructed, moving beyond the constraints of standard block-wise fully connected layers. At the core of this discovery is the MTS factorization, a structured operator designed to perform mode-wise linear transformations. Unlike previous deep learning-based compressive sensing (CS) methods that rely on local convolutional layers—often resulting in limited receptive fields—MTS-CSNet captures global and local features simultaneously. This multiscale approach ensures that the sampling operators can scale effectively for high-dimensional data, such as high-definition RGB images, without the exponential increase in computational overhead usually associated with dense networks. The researchers highlighted that traditional methods tend to struggle with high-dimensional signals because their sampling operators lack the necessary structural complexity to handle diverse data scales. By integrating tensor factorization directly into the neural architecture, the MTS-CSNet maintains efficiency while enhancing the quality of signal recovery. This methodological shift represents a significant step forward in the field of computational imaging, potentially impacting how digital cameras and medical imaging devices capture and process large datasets under bandwidth constraints.