Comparative Analysis of Pre-Trained Deep Learning Architectures in the Study of Multi-Instance Multi-Label Dataset

Multi-instance multi-label (MIML) learning deals with instances having multiple attributes simultaneously. Lael prediction for natural scene image prediction is vital for recognition, detection, and understanding, enabling machine comprehension for diverse applications. In the proposed work, three transfer learning architectures namely residual network (ResNet152V2), mobile network (MobileNetV2), and inception network (InceptionV3) are employed for the automated classification of natural scene images. The work offers a comparative analysis based on various evaluation metrics such as area-under-curve (AUC), F1 Score, accuracy, and loss. The evaluation includes a comprehensive analysis of the dataset and interpretability plots, incorporating techniques like grad class activation maps (Grad CAM), shapley additive explanations (SHAP), local interpretable model-agnostic explanations (LIME), saliency maps, and smooth grad CAMs. Among these transfer learning models, ResNet-152V2 demonstrated the highest precision, achieving AUC, F1- score, accuracy, and loss of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$97.21 \%, 83.54 \%, 82.01 \%$</tex>, and 0.3642, respectively, on the validation data in comparison to the other two models.