Image Frame Error Detection Using Candidate Regions and Object Checks

The patent divides the error detection task into two distinct stages performed by separate machine learning models: (1) a first ML model identifies candidate error areas in the image frame, and (2) a second ML model performs object detection to verify whether objects are present in those candidate areas. This segmentation allows each model to specialize in specific aspects of error detection, improving overall confidence while maintaining manageable system complexity through modular architecture.

The patent introduces an intermediary verification step using object detection as a mediator between initial error candidate identification and final error confirmation. The object detection model acts as a secondary validator that checks whether the identified candidate error areas contain actual objects, thereby increasing detection confidence without requiring a complete redesign of the detection pipeline.

The first machine learning model performs preliminary action by quickly identifying candidate error areas in the image frame before the more computationally intensive object detection is performed. This preliminary filtering step reduces the scope of subsequent verification by focusing object detection only on suspected error regions rather than the entire image, thereby minimizing additional processing time while improving reliability.

The patent applies partial action by performing object detection only in candidate error areas identified by the first model, rather than conducting full-image object detection. This selective verification approach applies the additional computational effort only where needed (in candidate error regions), reducing overall processing time while maintaining improved detection reliability through targeted verification.