Cognitively robust model using a deep learning framework for target recognition under suboptimal illumination in intelligent video surveillance system

This paper presents a cognitively robust deep learning framework for target recognition under suboptimal illumination, which is developed to enhance the recognition performance and stability of intelligent video surveillance systems. The proposed framework integrates three complementary modules: an Illumination-Adaptive Enhancement Module (IAEM) that performs LIME-based illumination estimation and correction followed by linear contrast adjustment; a Deep Recognition Backbone (DRB) based on transfer-learned CNNs for robust feature extraction; and a Cognitive Robustness Evaluator (CRE) that quantitatively assesses recognition stability across optimization and illumination conditions. Importantly, CRE operates exclusively as a post-training evaluation module and does not participate in model training, parameter optimization, loss computation, or inference. The framework was evaluated on the Extended Yale Face Database B and NIR Dataset for closed-set face recognition and on the BDD100K Night Subset for multi-class object detection. The proposed IAEM–DRB pipeline achieved classification accuracies of 99.81% and 99.62% on the two face-recognition datasets, respectively, while BDD100K performance was evaluated using detection-specific metrics. Under illumination levels ranging from 20 to 400 lx, recognition accuracy varied from 98.16% to 99.87%, with a Cognitive Robustness Index (R c ) of 0.981. Component-wise ablation experiments further examined the effects of input enhancement, backbone selection, and alternative enhancement methods. The results demonstrate high recognition accuracy and limited performance variation under the evaluated illumination conditions. The results demonstrate high recognition accuracy and limited performance variation under the evaluated illumination conditions, therefore, the framework showing its potential for illumination-robust video surveillance systems.