Fully convolutional network

As explained in Inside Deep Learning: Math, Algorithms, Models

A neural network composed only of convolutional layers plus supporting layers such as nonlinearity and normalization, with no dense linear layers.

A fully convolutional network contains convolutional layers throughout, together with supporting nonlinearity and normalization layers, but no dense linear layers. Using convolution for the output preserves local image structure and produces predictions across an entire image, making the architecture suitable for image segmentation. section 8.1

A practical segmentation pipeline uses the network with image-shaped labels:

• Resize images to 256 × 256, remove alpha channels, move channels to PyTorch’s first dimension, and normalize pixel values. section 8.1.1
• Merge separate nucleus masks with a logical OR, producing one binary mask whose pixels indicate nucleus presence. section 8.1.1
• Train the network to predict the complete mask rather than a single image-level or object-level output. section 8.1 section 8.1.1

Figure 8.2 This is how we process the Data Science Bowl data for image segmentation. Each folder under the root directory has several subfolders: an images folder with one image (I know, confusing) and a masks folder that has a binary mask for each nucleus. The masks are separated for object detection (we’ll get to that later), so we merge them all into one image that indicates where masks are (class 1) and are not (class 0).

Follow the image-and-mask transformation from separate nucleus masks to one binary segmentation target. section 8.1.1

How Inside Deep Learning: Math, Algorithms, Models covers it

How deep it goes

Defines the architecture and applies it to preprocessing and whole-image prediction, but does not detail its layer-by-layer implementation.

Examples

Uses cell-microscopy images, merged nucleus masks, and a 256 × 256 Data Science Bowl dataset pipeline.

What you can do after

Apply it via PyTorch to specify a network that predicts a complete segmentation mask from an input image.

How it connects

Connects convolutional neural networks, image segmentation, normalization, nonlinearity, and object detection.

liveBookmanning.com

Read Inside Deep Learning: Math, Algorithms, Models on liveBook Get the book at manning.com

Related concepts

• generationuses · Generation is the computational production of outputs from supplied inputs or context, often by applying a model to produce predictions or new content.
• linear layercontrasts with · A linear layer is a neural-network layer that maps input features to outputs by a learned affine transformation, typically connecting each output to every input feature.
• convolutional neural networkkind of · A convolutional neural network is a neural-network architecture that applies learned convolutional operations to extract patterns from structured data.
• Region Proposal Networkequivalent to · A Region Proposal Network is a neural-network module that generates candidate object regions from image features for a two-stage detector.

This explainer was generated by Manning's knowledge graph from the text of Inside Deep Learning: Math, Algorithms, Models. The section links point to the parts of the book it draws on.