Feature Pyramid Network (FPN)

A Feature Pyramid Network is a neural networkNeural NetworkA Neural Network is a computing system inspired by the structure of biological neural networks in the brain. It forms... architecture that constructs a multi-scale feature pyramid from a single input image by combining bottom-up feature extraction with top-down feature propagation and lateral connections. The bottom-up pathway extracts features at progressively lower resolutions through a backbone network like ResNetResNetResNet (Residual Network) is a deep neural network architecture that uses skip connections to enable training of very.... The top-down pathway upsamples semantically strong but spatially coarse features from higher pyramid levels. Lateral connections merge these upsampled features with corresponding bottom-up features at each scale, producing feature maps that are both semantically rich and spatially precise at all resolutions. Multi-Scale Detection in Geospatial ImageryFPN is critical for geospatial applications because objects in satellite and aerial imageryAerial ImageryAerial imagery involves photographs taken from planes or drones, offering detailed views of Earth's surface. It is a ... span enormous size ranges. A single image may contain small objects like vehicles at just a few pixels across alongside large structures like industrial complexes spanning thousands of pixels. Without multi-scale features, detectors struggle with this size variation. FPN enables consistent detection accuracy across scales by providing appropriate feature representations for objects of every size. Building detection systems use FPN to find everything from small residential structures to large commercial buildings. Infrastructure monitoring uses FPN-based detectors to identify features ranging from individual utility poles to entire solar farms. Integration and EvolutionFPN is typically integrated with object detectionObject DetectionObject Detection is a computer vision technique that identifies and localizes specific objects within images or video... frameworks like Faster R-CNN, RetinaNet, and Mask R-CNN, providing the feature backbone that these detectors build upon. It also enhances semantic segmentationSemantic SegmentationSemantic Segmentation is a computer vision technique that assigns a class label to every pixel in an image, enabling ... by providing multi-scale context for pixel classification. Modern extensions include PANet (Path Aggregation Network) that adds a bottom-up path to the pyramid, BiFPN (Bidirectional FPN) from EfficientDet that learns weighted feature fusion, and NAS-FPN that uses neural architecture search to optimize the pyramid structure.