Breaking Tradition With Perception: Debiasing Strategies in Cloth‐Changing Person Re‐Identification

IET Image Processing, Journal Year: 2025, Volume and Issue: 19(1)

Published: Jan. 1, 2025

ABSTRACT Person ReID aims to match images of individuals captured from different camera views for identity retrieval. Traditional methods primarily rely on clothing features, assuming that do not change clothes in a short time frame. This assumption significantly reduces recognition accuracy when changes, particularly long‐term tasks cloth‐changing person re‐identification (CC‐ReID). Thus, achieving effective clothing‐change scenarios has become critical challenge. paper proposes an automatic perception model (APM) address the break posed by changes. The uses dual‐branch with dynamic learning (DPL) strategy and branch, minimizing bias introduced while preserving semantic features. DPL dynamically adjusts training weights enhance model's ability learn varying sample difficulties feature distributions. branch captures deeper relationships, alleviating impact improving distinguish intra‐class transformations. Validated Celeb‐Reid Celeb‐Reid‐light datasets, APM achieves mean average precision (mAP) 22.6% 25.9%, Rank‐1 77.3% 79.5%, respectively. It also excels short‐term ReID, 90% mAP 96.3% Markt1501, demonstrating robustness across scenarios.

Language: Английский

Advancing Glaucoma Diagnosis Through Multi‐Scale Feature Extraction and Cross‐Attention Mechanisms in Optical Coherence Tomography Images

Engineering Reports, Journal Year: 2025, Volume and Issue: 7(4)

Published: April 1, 2025

ABSTRACT Glaucoma is a major cause of irreversible vision loss, resulting from damage to the optic nerve. Hence, early diagnosis this disease crucial. This study utilizes optical coherence tomography (OCT) images “Shahroud Eye Cohort Study” dataset which has an unbalanced nature, diagnose disease. To address imbalance, novel approach proposed, combining weighted bagging ensemble learning with deep models and data augmentation. Specifically, glaucoma expanded sixfold using augmentation techniques, normal stratified into five groups. samples were subsequently merged each group, independent training was performed. In addition balancing, proposed method incorporates key architectural innovations, including multi‐scale feature extraction, cross‐attention mechanism, Channel Spatial Attention Module (CSAM), improve extraction focus on critical image regions. The suggested achieves impressive accuracy 98.90% 95% confidence interval (96.76%, 100%) for detection. comparison earlier leading methods ConvNeXtLarge model, our exhibits 2.2% improvement in while fewer parameters. These results have potential significantly aid ophthalmologists detection, more effective treatment interventions.

Language: Английский