该文件“CVPR2019-ocr.zip”可能包含与计算机视觉和光学字符识别相关的资源。

简介：
Online handwritten Chinese text recognition (OHCTR) presents a considerable challenge due to the extensive character set involved, the inherent ambiguity in segmenting the text, and the variability of input sequences. This research leverages the exceptional proficiency of path signatures to transform online pen-tip movements into insightful signature feature maps, employing a sliding window approach. This method effectively captures both the analytical and geometric characteristics of pen strokes, exhibiting a notable degree of local invariance and resilience. To address this, we introduce a multi-spatial-context fully convolutional recurrent network (MC-FCRN), designed to harness the diverse spatial contexts present within these signature feature maps and subsequently generate a prediction sequence—all while circumventing the complexities associated with traditional segmentation techniques. Moreover, we have developed an implicit language model that enables predictions based on semantic context within a sequence of predicting features, thereby offering a novel approach for integrating lexicon constraints and prior knowledge regarding a specific language into the recognition process. Evaluations conducted on two established datasets – Dataset-CASIA and Dataset-ICDAR – demonstrated remarkably strong performance, achieving accuracy rates of 97.10% and 97.15%, respectively. These results represent a substantial improvement over previously reported state-of-the-art performance in the field.