<![CDATA[Image-based distracted driver classification has been widely studied to reduce the risk of traffic accidents, yet most studies use data captured under uniform lighting conditions. However, illumination variation can degrade feature extraction in Convolutional Neural Network (CNN) architectures. Channel attention (CA) has previously been shown to improve EfficientNetB0, but EfficientNetB0 with CA has not yet been extensively tested for distracted driver classification under diverse lighting conditions. This study aims to evaluate the performance of channel attention on EfficientNetB0 in classifying distracted driver activities under diverse lighting conditions. We used the State Farm Distracted Driver Dataset and a primary dataset of 10.080 images from three subjects captured under dark, semi-bright, and bright lighting conditions. The secondary data and two primary subjects were used for training, while one unseen primary subject was used for testing. Preprocessing included image resizing, normalization, and data augmentation. The models were trained using two-phase transfer learning (frozen backbone followed by fine-tuning) and evaluated using accuracy, precision, recall, and F1-score. The accuracies of baseline EfficientNetB0 versus EfficientNetB0+CA were 38.84% vs 60.36% (dark), 61.07% vs 32.59% (semi-bright), 67.14% vs 84.73% (bright), and 55.68% vs 59.23% (all conditions). CA improved accuracy in dark and bright conditions but reduced it in semi-bright conditions, so a consistent performance gain from adding CA cannot be concluded.]]>
