cs.CVOct 27, 2026

CVBL-CCTV-Face Dataset for Real-World Face Enhancement and Recognition in Surveillance

Authors: Manish Das, Ajay Kumar Yadav, Shiv Ram Dubey, Satish Kumar Singh

Organizations: Computer Vision and Biometrics Lab, Indian Institute of Information Technology, Allahabad

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

In real-world surveillance settings, facial captures from closed-circuit television systems suffer from degradations such as low resolution, motion blur, noise, and occlusions, which undermine dependable face recognition. In this study, we propose a framework for super-resolution and identity recognition, aimed at surveillance applications. For this purpose, we present the CVBL-CCTV-Face Dataset which is a dataset of real-world images taken in an unpaired manner where each pair consists of a low-quality (LQ) image of a face captured by a 8MP IP Dome CCTV camera together with a high-quality (HQ) reference image that is taken in controlled settings. In order to cope with the various problems posed by such datasets, we construct a modular system that involves face detection, deep learning-based super-resolution of low-quality face images, and identity recognition through L2-normalized embeddings extracted by an embedding model. Enhancement models are trained and tested according to three different setups, and the system is deployed for identity verification in a real-time unconstrained environment.

Figures & tables

Fig. 1: Overview of the proposed face enhancement and recognition pipeline integrating face detection, enhancement, and identity recognition.
Fig. 2: Training and evaluation configurations: (i) Pretrained models, (ii) Fine-tuning on synthetic HQ degradation, (iii) Fine-tuning on real LQ samples. Emb Extractor denotes the embedding extraction model.
Fig. 3: Real-time surveillance system with live CCTV stream processing, on-the-fly enhancement, and identity recognition.
Fig. 4: CVBL-CCTV-Face Dataset : top row shows high-quality (HQ) references, bottom row shows corresponding low-quality (LQ) images.
Fig. 5: Examples of typical challenges in the CVBL-CCTV-Face Dataset . Row 1: Blurred Faces due to motion and poor focus. Row 2: Faces with Caps causing occlusion. Row 3: Partial Faces obstructed by doors or other objects. Row 4: Side Faces with significant pose variation. Row 5: Head-Down Faces.
Fig. 6: Proposed pipeline: (a) MTCNN face detection; (b) Super-resolution via facial enhancement models; (c) Inception-ResNet-v1 embedding extraction and L2 norm identity matching (M × 512 SR, P × 512 HR embeddings).
Fig. 7: Illustration of various challenges encountered by MTCNN during face detection, highlighting the challenges in proposed CVBL-CCTV-Face Dataset .
Fig. 8: Face enhancement examples using pretrained CodeFormer [ 25 ] , RestoreFormer++ [ 26 ] , PiSA-SR [ 19 ] , and TSD-SR [ 20 ] (left to right).
Fig. 9: Visual comparison between LQ/HQ inputs and restored outputs under different training strategies. Each column shows outputs from CodeFormer, TSD-SR, and RestoreFormer++ models for Pretrained, Fine-tuned on synthetic degradation, and Fine-tuned on real LQ images.
Fig. 10: Real-time results: Left shows enrolled HR faces; right shows enhanced live detections with matched identities.
Fig. 11: Facial enhancement and identity recognition results from the CVBL-CCTV-Face Dataset . Incorrect identity matches are highlighted with red boxes and their corrected matches after enhancement are shown with green boxes.
Fig. 12: Qualitative face enhancement results on four identities from the CVBL-Face dataset . Each row corresponds to a different individual. From left to right: (1) original low-quality (LQ) surveillance face crop, followed by enhanced outputs from (2) CodeFormer [ 25 ] , (3) RestoreFormer++ [ 26 ] , (4) PiSA-SR [ 19 ] , and (5) TSD-SR [ 20 ] .
ModelTraining Setting
 Pre-TrainedFT (Real LQ)FT (Synthetic)
PiSA-SR [ 19 ]64.773.2367.81
TSD-SR [ 20 ]64.4533.3370.08
CodeFormer [ 25 ]57.6940.2059.20
RestoreFormer++ [ 26 ]59.4662.7567.49
a Baseline (No enhancement): 62.60%.
TABLE I: Benchmark Comparison of Representative Enhancement Methods Using Top-1 Recognition Accuracy (%)
More in this topic

Face Recognition

Face recognition matches or identifies people from facial images, videos, and biometric representations across changing conditions.

See all 108 papers

Explore similar work

CardsList

  1. Robust Face Super-Resolution and Recognition Through Multi-Feature Aggregation in Diffusion Models

    Jul 6, 2026Marcelo dos Santos, Rayson Laroca, João Carlos Raposo Neves +1Diffusion Models for Image RestorationFace Recognition

    Federal University of Paraná, Department of Informatics, Curitiba-PR, 81531-970, Brazil · Pontifical Catholic University of Paraná, Federal University of Paraná · University of Beira Interior +1

  2. FANVIDv2: Evaluating Video Super-Resolution by Face and Licence-Plate Recognition Under Compound Degradation

    Sep 30, 2026Kavitha Viswanathan, Vrinda Goel, Shlesh Gholap +5Video Super-Resolution

    Department of Electrical Engineering, Indian Institute of Technology Bombay, Mumbai 400076, India

  3. Collaborative Feature Aggregation for Face Super-Resolution and Robust Re-Identification

    Jul 30, 2026Juheon Hwang, Taewan Kim, Jiwoo KangImage Super-ResolutionPerson Re-Identification

    Department of Electrical and Electronic Engineering, Yonsei University, Seoul, 03722, South Korea. · Data Science Major, Dongduk Women’s University, Seoul, 02748, South Korea. · Division of Artificial Intelligence Engineering, Sookmyung Women’s University, Seoul, 04310, South Korea.