A Region Proposal Network (RPN) is a deep learning component used in object detection models to efficiently generate candidate object regions (also called proposals) in an image. It was introduced as part of the Faster R-CNN architecture and significantly improved detection speed by replacing slow, external proposal methods (like Selective Search) with a trainable, end-to-end module.

The RPN slides a small convolutional network over the feature map of the input image and, at each location, predicts:

• Objectness score: the likelihood that the region contains an object,
• Bounding box coordinates: adjustments to predefined anchors (reference boxes of various scales and aspect ratios).

These predicted boxes serve as proposals for where objects might exist and are passed on to subsequent stages (e.g. classification and refinement). Because the RPN shares features with the detection network, it adds minimal overhead and allows for joint training with the detector.