background bias

**Background Bias** is the **tendency of image classifiers to rely on background context for classification instead of the actual object** — the model learns to associate specific backgrounds with specific classes (e.g., boats with water, cows with grass), failing when objects appear in unusual contexts. **Background Bias Examples** - **Context Association**: "Cow" = "green background" — model classifies any green-background image as containing a cow. - **Outdoor/Indoor**: Class predictions correlate with indoor/outdoor background rather than the object. - **Inpainting Test**: Replace the background with a random background — accuracy drops significantly for biased models. - **Foreground Test**: Show only the object (no background) — biased models lose significant accuracy. **Why It Matters** - **False Correlation**: Background features correlate with labels in training data but are not causally related. - **Deployment**: In real-world deployment, objects appear in diverse backgrounds — background-biased models fail. - **Semiconductor**: Defect classifiers may learn imaging system artifacts (background patterns) instead of actual defect features. **Background Bias** is **reading the wallpaper instead of the book** — classifying based on background context rather than the actual object of interest.