Details

Generating Photorealistic Facial Composites with Diffusion Models
ID Miočić, Matej (Author), ID Peer, Peter (Mentor) More about this mentor... This link opens in a new window, ID Tomašević, Darian (Comentor)

.pdfPDF - Presentation file, Download (22,44 MB)
MD5: CA1209EA29A1371C8E6E8572B3FBFB2B

Abstract
Facial composites, also known as police sketches, are essential tools in law enforcement for reconstructing a suspect’s appearance based on eye witness descriptions. Traditionally, forensic artists manually create these sketches by working closely with eyewitnesses, which is often slow and te dious. To simplify this process, modern law enforcement increasingly relies on advanced computer software. While deep learning has introduced several innovative approaches in recent years, the use of diffusion models for this task remains largely unexplored. We present Diff-FIT (Diffusion Facial Identification Technique), a novel framework for generating photorealistic facial composites using diffusion models. Diff-FIT enables fast generation of initial images from a textual description, followed by intuitive, sequen tial edits based on ongoing eyewitness input. Our approach generates facial composites with a preference comparable to existing methods, while enabling a greater range of facial variation and more diverse adjustments, without compromising identification performance or image quality. In a user study involving biometric experts and non-experts, facial composites generated by Diff-FIT were rated on par with those from state-of-the-art methods in both subjective evaluations and identification rates.

Language:English
Keywords:Computer Vision, Deep Learning, Image-Based Biometrics, Diffusion Models, Facial Composite
Work type:Master's thesis/paper
Typology:2.09 - Master's Thesis
Organization:FRI - Faculty of Computer and Information Science
Year:2025
PID:20.500.12556/RUL-171779 This link opens in a new window
COBISS.SI-ID:248392963 This link opens in a new window
Publication date in RUL:02.09.2025
Views:325
Downloads:101
Metadata:XML DC-XML DC-RDF
:
Copy citation
Share:Bookmark and Share

Secondary language

Language:Slovenian
Title:Generiranje fotorealističnih fotorobotov z difuzijskimi modeli
Abstract:
Fotoroboti, znani tudi kot rekonstrukcije iskanih oseb, so ključno orodje pri identifikaciji osumljencev na podlagi opisov očividcev. V preteklosti so te rekonstrukcije ročno izdelovali forenzični risarji v sodelovanju z očividci, kar je bil dolgotrajen in zahteven postopek. Čeprav je globoko učenje že omogočilo številne nove pristope, je uporaba difuzijskih modelov za ta namen še vedno slabo raziskana. V tem magistrskem delu predstavimo Diff-FIT (Diffusion Facial Identification Technique), novo ogrodje za ustvarjanje fotorealističnih fotorobotov s pomočjo difuzijskih modelov. Diff-FIT omogoča hitro generiranje začetnih fotorobotov iz besedilnega opisa, ki mu sledijo intuitivni nadaljnji popravki na podlagi sprotnih opisov očividcev. Naš pristop omogoča izdelavo fotorobotov s primerljivo preferenco kot obstoječe metode, hkrati pa omogoča širši razpon obraznih variacij in bolj raznolike prilagoditve brez poslabšanja prepoznavnosti ali kakovosti slike. V uporabniški študiji, v katero so bili vključeni biometrični strokovnjaki in nestrokovnjaki, so fotorobote, izdelane z Diff-FIT, ocenili primerljivo s fotoroboti, izdelanimi z najsodobnejšimi metodami, tako v subjektivnih ocenah kot v uspešnosti identifikacije.

Keywords:računalniški vid, globoko učenje, slikovna biometrija, difuzijski modeli, fotorobot

Similar documents

Similar works from RUL:
Similar works from other Slovenian collections:

Back