We propose a general learning based framework for solving nonsmooth and nonconvex inverse problems with application to low-dose CT (LDCT) reconstruction. We model the regularization function as the combination of a sparsity enhancing and a non-local smoothing regularization. We develop an efficient learned descent-type algorithm (ELDA) to solve the nonsmooth nonconvex minimization problem by leveraging the Nesterov’s smoothing technique and incorporating the residual learning structure. We proved the convergence of the algorithm and generate the network, whose architecture follows the algorithm exactly. Our method is versatile as one can employ various modern network structures into the regularization, and the resulting network inherits the convergence properties, and hence is interpretable. We also show that the proposed network is parameter-efficient and its performance compares favorably to the state-of-the-art methods.

---

https://kaist.zoom.us/j/82680768716?pwd=4jDj5hW70PKYbTcYq1nbkEa9Gsarhi.1 Meeting ID: 826 8076 8716 Passcode: 933841 참고: Jan 16, 2025 07:00 PM Eastern Time (US and Canada) https://kaist.zoom.us/j/82680768716?pwd=4jDj5hW70PKYbTcYq1nbkEa9Gsarhi.1 Meeting ID: 826 8076 8716 Passcode: 933841