A Deep Learning-Based Image Processing Framework for Automated Feature Extraction, Classification, and Intelligent Decision Support

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Deep learning has strengthened automated image analysis by enabling hierarchical feature extraction and classification without extensive handcrafted feature engineering. This study developed an integrated image processing framework combining deep feature learning, multiclass classification, explainability, and confidence-aware decision support. A balanced dataset of 4,000 images across four classes was divided into training (70%), validation (15%), and testing (15%) subsets. A custom convolutional neural network (CNN), ResNet50, DenseNet121, and EfficientNetB0 were evaluated using accuracy, precision, recall, F1-score, and area under the curve (AUC). Grad-CAM was incorporated for visual interpretation, while prediction probabilities were stratified into confidence levels for decision support. EfficientNetB0 achieved the strongest performance, with 94.17% accuracy, 0.943 precision, 0.942 recall, 0.942 F1-score, and 0.982 AUC, exceeding the other evaluated architectures. The model maintained a macro-average F1-score of 0.942 across classes. High-confidence predictions achieved 97.8% accuracy, compared with 90.1% and 72.7% for intermediate- and low-confidence predictions, respectively. These findings demonstrate the potential of combining transfer learning, explainability, and confidence assessment to support accurate and interpretable image-based decision processes.