Fig. 3: Grad-CAM visualization of feature extraction by pretrained CNN and ViT models on “Good-quality”, “Under-extrusion”, and “Over-extrusion” images.
From: Transformer-based approach for printing quality recognition in fused filament fabrication
The heatmaps represent the focus intensity of each model, with red indicating high focus and blue indicating low focus. ViT heatmaps effectively highlight critical areas, such as under-extrusion regions along boundaries and over-extrusion blobs in the filming region, demonstrating the model’s ability to focus on relevant defect features.
