Generalized Imaging Augmentation via Linear Optimization of Neurons

Most deep networks for computational photography tasks require large-scale training, which is time-consuming, computing-cost, and even hard to implement for certain data-unaccessible tasks. The emerging untrained convolutional networks (CNNs) rely on explicit physical models whose discrepancies and disturbances would lead to unsatisfactory performance. In response to these challenges, this work reports a generalized augmenting technique for computational photography techniques based on LInear Optimization of Neurons (LION). LION linearly transforms the neurons of a pre-trained CNN and optimizes the transformation coefficients using a model-free color and texture regularization. Leveraging the inherent representation capabilities of the deep feature domain, we can enhance the quality of output images through a simple linear transformation of the pre-trained network features, without modifying network parameters or architecture. Furthermore, inspired by the concept of deep image prior, we develop a generalized workflow based on LION for augmenting untrained networks and conventional methods. A series of experiments have validated the technique’s effectiveness for general imaging augmentation in underwater, low-light, and computational lensless imaging applications.