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

I. INTRODUCTION

• The dataset addresses a benchmarking gap: prior face-recognition benchmarks largely use studio or synthetically degraded images, not CCTV faces with multiple uncontrolled degradations.The cited degradations include low resolution, motion blur, sensor noise, partial occlusion, and wide pose variation.Source: §I · p. 1
• CVBL-CCTV-Face provides 223 identities with authentic CCTV-degraded LQ images and separately acquired, spatially unaligned HQ references.Unlike datasets using synthetic degradation, it captures compound real-world distortions.Source: §I · p. 1
• The benchmark pipeline combines MTCNN detection, four transformer/diffusion enhancement models, and L2 embedding-based recognition, with a real-time CCTV deployment.The deployed system processes live streams and overlays bounding boxes and identity labels.Source: §I · p. 1
• Three configurations evaluate pretrained models, fine-tuning on synthetically degraded HQ images, and fine-tuning on real LQ samples.The configurations are used to characterize dataset challenges for surveillance performance.Source: §I · p. 1

II. RELATED WORK

• Earlier super-resolution methods optimize pixel fidelity on bicubic-downsampled LQ–HQ pairs, while authentic CCTV distortions are heterogeneous.Later approaches model varied degradations or improve perceptual quality, but GAN-based losses can introduce artifacts.Source: §II · p. 2
• Diffusion and newer architectures target blind super-resolution, recognition utility, or efficiency, while specialized face-enhancement models use component-aware priors, vector quantization, codebooks, and key–value memory.The cited discussion says these models mainly target expression analysis and reidentification in relatively controlled environments using synthetic degradations.Source: §II · p. 2
• Dataset examples span motion- or focus-blurred faces, caps, partial obstruction, side poses, and downward head tilts.The related-work discussion concludes that synthetic degradation limits real-life application and motivates surveillance-specific datasets and techniques.Source: Figure 5 · p. 2, §II · p. 2

A. CVBL-CCTV-Face Dataset

• The dataset records LQ faces with an 8MP IP dome camera in unconstrained indoor settings and captures HQ references independently under controlled conditions.Version A has 102 identities with one HQ image each and 8,090 LQ images; Version B has 130 identities, 2,612 HQ images, and 7,572 LQ images.Source: A. CVBL-CCTV-Face Dataset · p. 2
• After removing nine shared identities, the combined dataset contains 223 identities and is split by identity into training and test sets.The test set has 130 identities, 1,581 LQ images, and 528 HQ images; the training set has 14,081 LQ images and 2,186 augmented HQ images.Source: A. CVBL-CCTV-Face Dataset · p. 2

B. Challenges in the Dataset

• Dataset challenges include motion or defocus blur, caps, obstruction by walls or doors, nonfrontal poses, and downward head tilts.These conditions obscure facial landmarks or features and complicate enhancement, restoration, or identification.Source: B. Challenges in the Dataset · p. 3
• The authors state that these varied degradations and poses motivate robust enhancement models for unconstrained surveillance conditions.Source: B. Challenges in the Dataset · p. 3

• The pipeline has three stages: face detection and alignment, super-resolution enhancement, and embedding extraction with identity matching.Source: A. Face Enhancement and Recognition… · p. 3
• Fig. 8 compares pretrained CodeFormer, RestoreFormer++, PiSA-SR, and TSD-SR enhancement outputs from left to right.Source: Figure 8 · p. 3

1) Face Detection and Alignment:

• MTCNN detects faces and localizes five landmarks for alignment, but realistic surveillance conditions produce missed faces and false positives.The cited failure examples include non-face regions such as hands and toes.Source: 1) Face Detection and Alignment · p. 3

2) Super-Resolution and Enhancement:

• Four enhancement models address surveillance degradations: CodeFormer, RestoreFormer++ [26], PiSA-SR [19], and TSD-SR [20].Their approaches include fidelity-weighted codebooks, memory-based prior retrieval, dual-LoRA optimization, and identity-score distillation, respectively.Source: 2) Super-Resolution and Enhancement · p. 3

3) Embedding Extraction and Recognition:

• Inception-ResNet-v1 extracts 512-dimensional ℓ2-normalized embeddings, and each super-resolved face is matched to the nearest high-quality identity by Euclidean distance.Accuracy is the proportion of correct matches.Source: 3) Embedding Extraction and Recognition · pp. 3–4

A. Training Configurations for Face Enhancement Models

• The evaluation uses MTCNN detection and alignment, VGGFace2-pretrained Inception-ResNet-v1 embeddings, and Top-1 identification accuracy.Experiments are conducted on the CVBL-CCTV-Face Dataset and assess enhancement and identity recognition.Source: A. Training Configurations for Face… · p. 4

1) Setting 1: Pretrained Models:

• Pretrained enhancement models are applied without fine-tuning; they improve over raw low-quality inputs but often miss domain-specific characteristics.The passage notes reduced robustness in real-world surveillance scenarios.Source: 1) Setting 1: Pretrained Models · p. 4

2) Setting 2: Fine-Tuning on Synthetically Degraded HQ Images:

• Fine-tuning on synthetically degraded high-quality images uses Gaussian blur, added noise, and compression artifacts to emulate surveillance degradations, and results in substantial recognition-accuracy improvements.Source: 2) Setting 2: Fine-Tuning on… · p. 4

3) Setting 3: Fine-Tuning on Real LQ Images:

• Direct fine-tuning on real low-quality surveillance images did not improve over the baseline.The images include occlusions, pose variations, motion blur, and lighting changes.Source: 3) Setting 3: Fine-Tuning on Real LQ… · p. 4

B. Real-Time Enhancement and Recognition System

• Fig. 9 compares outputs from pretrained, synthetic-degradation fine-tuned, and real-LQ fine-tuned models across CodeFormer, TSD-SR, and RestoreFormer++.Source: Figure 9 · p. 4
• Fig. 10 shows enrolled high-resolution faces alongside enhanced live detections with matched identities.Source: Figure 10 · p. 4
• The live system detects faces with MTCNN, enhances them, compares Inception-ResNet-v1 embeddings with a high-resolution reference gallery, and displays the closest match with a bounding box and identity label.Source: B. Real-Time Enhancement and… · p. 4

A. Face Enhancement and Recognition Performance

• ID 32 was corrected to ID 108 after enhancement, with L2 distance changing from 0.9332 to 0.7562; an occluded face was also recovered as ID 100.The displayed examples include incorrect matches in red and corrected matches in green.Source: A. Face Enhancement and Recognition… · p. 4, Figure 11 · p. 5
• TSD-SR and RestoreFormer++ preserved identities most effectively across four identities in the qualitative comparison.The comparison reports L2 distances for CodeFormer, PiSA-SR, RestoreFormer++, and TSD-SR for each identity.Source: A. Face Enhancement and Recognition… · p. 5, Figure 12 · p. 5

B. Real-Time Face Enhancement and Recognition Results

• Approximately 25 FPS was achieved on an NVIDIA GeForce RTX 3090 under diverse CCTV conditions, including motion blur, occlusion, illumination changes, and non-frontal poses.The live system enhances detected faces and matches their embeddings to an HR gallery; Person2 was an incorrect match.Source: B. Real-Time Face Enhancement and… · p. 5
• The real-time results support enhancement as a preprocessing step for recognition under surveillance degradations, although ambiguous low-quality inputs can still produce incorrect matches.The example of Person2 is identified as an incorrect match.Source: B. Real-Time Face Enhancement and… · p. 5

C. Evaluation Under Different Training Configurations

• The evaluation pipeline uses MTCNN face detection, transformer- and diffusion-based enhancement, and Inception-ResNet-v1 feature extraction; performance is measured by L2 distance and top-1 identification accuracy.The feature extractor is trained on VGGFace2.Source: C. Evaluation Under Different Training… · p. 5
• 64.77% for PiSA-SR and 64.45% for TSD-SR exceeded the 62.6% unenhanced baseline when using publicly pretrained models without fine-tuning.The passage characterizes these gains as slight and reports poor generalization to real CCTV degradations.Source: C. Evaluation Under Different Training… · p. 5

2) Fine Tuning on Real Low Quality Images:

• 62.75% for RestoreFormer++, 40.20% for CodeFormer, and 33.33% for TSD-SR were reported with different real-LQ fine-tuning approaches.The real LQ images are independently captured, spatially unaligned, and affected by multiple uncontrolled degradations, unlike synthetically degraded HQ images.Source: 2) Fine Tuning on Real Low Quality… · p. 5

3) Fine-Tuning on Synthetically Degraded HQ Images:

• 70.08% accuracy for TSD-SR, 67.81% for PiSA-SR, and 67.49% for RestoreFormer++ were reported after synthetic fine-tuning of HQ images.The conclusion states that training methods involving synthetic degradations produced the best recognition accuracy among those tested.Source: 3) Fine-Tuning on Synthetically… · p. 5, §VII · p. 6

VII. CONCLUSION

• The conclusion identifies unpredictable, complex real-life surveillance degradations as a current challenge for models, despite synthetic-degradation training producing the best recognition accuracy in these tests.It describes the real-time system as demonstrating the feasibility of the pipeline.Source: §VII · p. 6

VIII. FUTURE WORK

• Future work will add FAR/FRR verification metrics, inference latency, and computational cost measurements across hardware platforms.Source: §VIII · p. 6