Feature Likelihood Correction for Rare Pattern Detection

The patent applies local quality by differentiating the treatment of feature quantities based on their rarity. Common feature quantities use conventional weight calculation, while rare feature quantities (those appearing in fewer than a threshold number of learning images) undergo special correction to reduce their weight influence. This localized differentiation resolves the contradiction by allowing high weights for difficult common features while suppressing potentially erroneous high weights for rare features.

The patent changes the parameter of feature quantity weight by introducing a correction mechanism. When a feature quantity is identified as rare, its weight is corrected using a formula that incorporates the count of learning images containing that feature. This parameter transformation converts potentially harmful high weights into appropriately scaled weights, improving detection accuracy for rare feature patterns while maintaining reliability for common features.

The patent applies preliminary anti-action by proactively identifying rare feature quantities before they cause detection instability. The system pre-calculates the count of learning images containing each feature quantity and prepares correction values in advance. When rare features are detected, the pre-prepared correction mechanism immediately counteracts their extreme weight influence, preventing detection instability before it occurs.

The patent implements feedback by using the count of learning images containing each feature quantity as a feedback signal. This feedback information is fed back into the weight calculation process, where it modulates the weight assignment. The feedback loop ensures that feature quantities with low occurrence counts (rare features) receive adjusted weights, creating a self-regulating system that maintains detection stability while preserving learning focus.