Saliency Maps for semiconductor inspection are visualizations that highlight which pixels in an image are most important for the model's output — computed by taking the gradient of the model's prediction with respect to the input image, revealing the sensitivity of the classification to each pixel.

Types of Saliency Maps

• Vanilla Gradient: $partial y / partial x$ — the raw gradient of the output with respect to input pixels.
• SmoothGrad: Average gradients over noisy versions of the input for less noisy maps.
• Integrated Gradients: Accumulate gradients along the path from a baseline to the input.
• Gradient × Input: Element-wise product of gradient and input for more visually interpretable maps.

Why It Matters

• Pixel-Level Explanation: Shows exactly which pixels influenced the classification at the finest granularity.
• Defect Localization: Saliency often highlights defect regions even without explicit localization training.
• Quality Assurance: Validates that inspection models respond to physical defect features, not imaging artifacts.

Saliency Maps are the pixel-level importance highlighter — showing which exact pixels drove the model's defect classification decision.