FCNs

Fully Convolutional Networks (FCNs) are a class of neural networks designed for dense prediction tasks in computer vision, in which all layers are convolutional, pooling, or other spatially preserving operations, and there are no fully connected layers that require fixed input sizes. FCNs produce spatially dense outputs such as per-pixel class probabilities, enabling semantic segmentation and related tasks.

They were introduced in 2015 by Long, Shelhamer and Darrell in the paper "Fully Convolutional Networks for Semantic Segmentation," which showed how a conventional CNN for image classification could be converted into a network that learns to classify each pixel. The key ideas include replacing fully connected layers with 1x1 convolutions to output class scores and using learned upsampling (deconvolution) to map low-resolution feature maps back to the input resolution. Skip connections from earlier layers help refine predictions by combining semantic information with high-resolution details.

Typical architecture features an encoder that extracts features with convolution and pooling, a classifier head comprised of 1x1 convolutions producing per-pixel scores, and a decoder that upsamples to the original image size. Variants such as FCN-8s, FCN-16s, and FCN-32s differ in how they combine information at multiple scales through skip connections.

FCNs have become foundational for semantic segmentation and have influenced many later models. They inspired architectures like U-Net, DeepLab, and PSPNet, which build upon the dense prediction concept with improved upsampling, atrous convolutions, and multi-scale context.

Limitations include computational and memory demands for high-resolution inputs and the need for large labeled datasets. Nevertheless, FCNs established a standard approach for end-to-end dense prediction in computer vision.