Authors:
K. Sadhana, T. Shreekumar
Addresses:
Department of Computer Applications, Mangalore Institute of Technology and Engineering, Moodabidri, Karnataka, India. Department of Computer Applications, Visvesvaraya Technological University (VTU), Belagavi, Karnataka, India. Department of Computer Science and Engineering, Mangalore Institute of Technology and Engineering, Moodabidri, Karnataka, India.
Abstract:
Face recognition systems often experience a significant drop in accuracy when images contain blur, noise, extreme pose, shadows, or occlusions. Conventional margin-based models treat all samples equally, thereby allowing degraded inputs to distort the embedding space during training. This work presents DAAM Net, a quality-aware face recognition framework that incorporates visual reliability directly into the margin learning process. The method combines a feature-extraction network with a lightweight quality-estimation module that assesses the clarity of each input image. This score is then used to generate a dynamic angular margin that strengthens discrimination for reliable samples and reduces the influence of those affected by degradation. A quality-regulated angular loss integrates these margins into the optimisation process, leading to more stable and compact identity representations. Experiments conducted on a curated subset of VGGFace2 show that DAAM Net achieves 97.6 per cent accuracy, outperforming fixed-margin methods such as ArcFace, CosFace, and SphereFace, with particularly strong gains on low-quality images. Ablation studies confirm the complementary value of explicit quality estimation and dynamic margin adjustment. The results indicate that incorporating visual quality into the training objective is an effective strategy for improving robustness in unconstrained face recognition scenarios.
Keywords: Face Recognition; Image Quality Variation; Dynamic Angular Margin; Quality Estimation; Deep Representation Learning; Robust Verification; Discriminative Embedding.
Received: 25/08/2024, Revised: 10/10/2024, Accepted: 05/12/2024, Published: 05/09/2025
DOI: 10.64091/ATICL.2025.000228
AVE Trends in Intelligent Computer Letters, 2025 Vol. 1 No. 3 , Pages: 132-152