Document Image Parsing and Understanding using Neuromorphic Architecture.

摘要

In this project, we investigate brain inspired information processing for text image recognition and its performance/accuracy optimization. The intelligent text recognition system (ITRS) works robustly on images with low quality by using a combination of input image data preparation, pattern matching and statistical inference. Our experimental results show that, compared to Tesseract, the ITRS achieves comparable accuracy for clean input images and higher accuracy for camera images with occlusions. Performance enhancement techniques are developed to reduce the processing speed at different layers. In the pattern matching layer, the computing power of multicore processors is explored to reduce the processing time. In the word confabulation layer, new data structures are adopted for the storage of a dictionary, which increases memory locality and reduces search complexity. In the sentence confabulation layer, different sentence models are compared and the best model with the highest accuracy is identified. Finally, the overall ITRS is implemented on a heterogeneous high performance computing cluster. It provides users with the flexibility of computing resource management through a configuration file.