Abstract: Image recognition and parsing techniques are provided herein. In the described examples, an input image, such as an image of a document (e.g., a scanned document), can be received. Scan mark candidates in the input image can be identified that correspond to blueprint scan marks for a stored set of form blueprints. The blueprint scan marks can indicate form entry areas or other features of a form associated with the form blueprint. Identified scan mark candidates can be compared with the corresponding blueprint scan marks. Based on the comparing, it can be determined that at least some of the scan mark candidates are confirmed scan marks. Based on the confirmed scan marks, one form blueprint can be identified that corresponds to the input image. Information can be extracted from the input image, for example by optical character recognition, based on the form blueprint to which the input image corresponds.

Abstract: Neural-network-based image segmentation techniques are provided herein. An input image that includes a plurality of characters can be received. Boundaries between the characters can be identified using a trained neural network. The input image can be segmented along the boundaries identified between the characters. The neural network can be trained using a training image and a training target vector. The training target vector can indicate one or more boundaries between characters in the training image. Neural-network-based segmentation can be used alone or in conjunction with other segmentation techniques to improve overall segmentation accuracy.