BEGIN:VCALENDAR VERSION:2.0 PRODID:-//wp-events-plugin.com//7.4.0.1//EN TZID:Asia/Kolkata X-WR-TIMEZONE:Asia/Kolkata BEGIN:VEVENT UID:222@cds.iisc.ac.in DTSTART;TZID=Asia/Kolkata:20260820T101500 DTEND;TZID=Asia/Kolkata:20260820T111500 DTSTAMP:20260817T050653Z URL:https://cds.iisc.ac.in/events/ph-d-thesis-colloquium-102-cds-20-august -2026-stable-denoiser-driven-regularization-from-kernel-methods-to-constra ined-and-pretrained-deep-models/ SUMMARY:Ph.D: Thesis Colloquium: 102: CDS: 20\, August 2026 "Stable Denoise r-Driven Regularization: From Kernel Methods to Constrained and Pretrained Deep Models" DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nPh.D. Thesis Col loquium\n\n\n\nSpeaker: Mr. Arghya Sinha\nS.R. Number: 06-18-01-10-12-21-1 -19979\nTitle: "Stable Denoiser-Driven Regularization: From Kernel Methods to Constrained and Pretrained Deep Models"\nResearch Supervisor: Prof. Ku nal N. Chaudhury and Prof. Debnath Pal\nDate &\; Time : August 20\, 202 6 (Thursday)\, 10:15 AM\nVenue : #102\, CDS Seminar Hall\n\n\n\nABSTRACT\n In classical model-based image reconstruction\, the forward measurement mo del is combined with an explicit regularizer that captures prior knowledge about the unknown image. Keeping these components separate allows the sam e regularizer to be paired with different forward models and used across i maging modalities. Regularizers promote meaningful spatial structure by su ppressing noise\, artifacts\, and unwanted variations. Since denoisers per form a closely related task\, they can naturally serve as regularizers. Th is connection underlies frameworks such as Plug-and-Play (PnP) and Regular ization by Denoising (RED)\, which incorporate an image denoiser into clas sical gradient and proximal reconstruction algorithms. These methods have proved highly effective in computational imaging problems such as deblurri ng\, superresolution\, tomography\, and magnetic resonance imaging.\n\nDes pite this success\, understanding when and why these methods converge rema ins an important open problem. Their analysis involves a fixed-point opera tor formed by the denoiser and the forward model. Classical convergence re sults require properties such as nonexpansiveness or averagedness. Modern deep denoisers are generally trained without such constraints\, since thes e can limit their expressive power and reconstruction performance. This cr eates a tradeoff between performance and theoretical guarantees. When an u nconstrained denoiser is applied repeatedly within PnP or RED\, the iterat es may fail to converge\, and reconstruction quality may initially improve but later deteriorate. This characteristic peak-and-collapse behavior mot ivates methods that retain the performance of modern deep models while mak ing iterative reconstruction stable and reliable.\n\nIn this thesis\, we s tudy this problem across a progression of increasingly expressive denoiser models. We begin with classical kernel denoisers\, which recover an image by averaging pixels or patches according to a prescribed similarity funct ion. Once the similarity weights are fixed\, the denoiser can be represent ed as a linear operator. This structure allows us to use tools from operat or and spectral theory to establish convergence\, uniqueness\, and linear- rate guarantees for several PnP and RED algorithms. We then carry these id eas into a trained setting. Instead of prescribing the similarity function \, we train it from data and use the resulting weights to construct a deno ising operator that remains linear for each input instance. By controlling its spectral properties\, we obtain trained denoisers that perform well e mpirically while remaining globally nonexpansive. Under suitable condition s on the forward model\, this makes the overall reconstruction operator co ntractive\, guaranteeing a unique fixed point and linear convergence of th e iterates.\n\nIn a different direction\, we ask whether powerful pretrain ed networks can be stabilized without retraining or modifying them. Treati ng these networks as black-box operators with no useful structural guarant ees\, we introduce a data-driven framework based on contractive anchoring. The method measures the local expansiveness of the reconstruction operato r through a stability index and adaptively blends its output with a lightw eight contractive anchor. Anchoring is used only when required\, allowing the original operator to remain dominant while keeping the iterates bounde d. The resulting method acts as a drop-in wrapper\, requires no additional parameter tuning\, and consistently suppresses peak-and-collapse behavior across computational imaging problems\, proximal algorithms\, and deep de noiser architectures.\n\nTogether\, the thesis shows how ideas from operat or theory and spectral theory can be used to understand structured kernel denoisers\, guide the training of stable denoisers\, and ultimately stabil ize powerful pretrained networks. This allows us to move from mathematical ly tractable models to modern trained denoisers while retaining meaningful stability guarantees for the resulting reconstruction algorithms.\n\n\n\n ALL ARE WELCOME CATEGORIES:Events,Ph.D. Thesis Colloquium END:VEVENT BEGIN:VTIMEZONE TZID:Asia/Kolkata X-LIC-LOCATION:Asia/Kolkata BEGIN:STANDARD DTSTART:20250820T101500 TZOFFSETFROM:+0530 TZOFFSETTO:+0530 TZNAME:IST END:STANDARD END:VTIMEZONE END:VCALENDAR