Non-max suppression refers to the idea of suppressing predicted information that is not predicted with the highest confidence.
Non-max suppression is commonly used in computer vision software, where a computer vision model may identify many nearby redundant edges or bounding boxes.
To reduce this redundancy, only the maximum feature in a set of related features are kept. For example, a computer vision model might identify a single object using multiple overlapping bounding boxes. The bounding box with the highest prediction probability is kept, and the overlapping boxes are removed in favor of the kept box. Other boxes may be unaffected by non-max suppression if they are too far away.
In the case of bounding boxes, non-max suppression is said to prune low-confidence bounding boxes that have a high intersection over union with the highest-confidence bounding box they intersect with.