QUALICLIPNET: A Hybrid Deep Learning Framework for Robust Image Quality Assessment and Adversarial Resilient Classification

This study presents a hybrid AI framework that integrates deep learning with adversarial training to enhance the quality assessment and colour prediction of images, such as onions. The process starts with the preprocessing and segmentation of the input images, which is then followed by feature extraction through a CLIP-based encoder framework. The derived embeddings are fed into a teacher model driven by GradeSimCNN and optimized using Knowledge Distillation (KD) to develop a more efficient student model. The fundamental ImageQualNet algorithm evaluates and organizes images according to their quality and conducts feature-based analysis for classification purposes. To enhance the resilience of the model, adversarial training is carried out using the Fast Gradient Sign Method (FGSM), which strengthens the framework against perturbed inputs by leveraging the Fast TRADE AI algorithm and also improves image classification by facilitating quick output colour predictions with Vision Transformers (ViTs). This multi-phase system guarantees precise classification and output prediction, even when facing adversarial noise, making it ideal for practical vision tasks that require both accuracy and robustness.