Classification with segmentation neural network for image-based content capture

摘要

A segmentation neural network is extended to provide classification at the segment level. An input image of a document is received and processed, utilizing a segmentation neural network, to detect pixels having a signature feature type. A signature heatmap of the input image can be generated based on the pixels in the input image having the signature feature type. The segmentation neural network is extended from here to further process the signature heatmap by morphing it to include noise surrounding an object of interest. This creates a signature region that can have no defined shape or size. The morphed heatmap acts as a mask so that each signature region or object in the input image can be detected as a segment. Based on this segment-level detection, the input image is classified. The classification result can be provided as feedback to a machine learning framework to refine training.