Image Processing Apparatus for Anomaly Detection

The patent divides the medical image into multiple local regions and generates normal images for each region separately, rather than generating one normal image from the entire input image. This segmentation approach allows each local normal image to capture region-specific characteristics while reducing the impact of individual differences across the whole image, thereby improving abnormal component separation accuracy.

The patent applies local quality by generating normal images with different synthesis strengths for different local regions. The synthesis strength is adjusted based on the degree of individual difference in each region, allowing regions with high individual differences to have lower synthesis strength (preserving more original information) while regions with low individual differences can have higher synthesis strength (achieving better normalization). This resolves the contradiction by optimizing the balance between structure information retention and anomaly detection accuracy for each local region.