Abstract
This paper describes the Address Block Location System (ABLS), an advanced framework designed to identify destination addresses on complex and unstructured mail pieces, such as magazines (flats) and irregular parcels. Utilizing a blackboard architecture, ABLS integrates over 20 specialized image-processing tools to generate and evaluate address candidates. By employing Dempster-Shafer theory for evidence combination and a utility-based tool selection process, the system achieved an 81% success rate on a challenging dataset, significantly outperforming the 55% industry standard of the late 1980s.
Introduction
In the late 1980s, automated mail sorting was highly effective for standard envelopes but struggled with "flats" and irregular parcels (IPPs) where address placement is often unpredictable. The core challenge was not optical character recognition (OCR) itself, but the initial localization of the address block amidst advertisements, labels, and graphics. This research presents ABLS, a system designed to handle environments ranging from highly structured letters to completely unstructured parcels by treating address location as a problem-solving task.
System Architecture and Methodology
The foundation of ABLS is a Blackboard Architecture, a classic AI model where multiple specialized "tools" or knowledge sources contribute evidence to a central data structure (the blackboard). This allows the system to operate opportunistically, switching between bottom-up processing (grouping pixels into characters and lines) and top-down processing (using high-level features like return addresses to narrow the search area).
The Tool Box and Dependency Graph
The system utilizes a suite of more than 20 specialized routines categorized into candidate generation, discrimination, and control mechanism tools. To optimize performance, a directed dependency graph dictates the sequence of tool applications, ensuring that a tool is only considered when its necessary inputs are available.
Tools are selected based on a "utility value" that accounts for historical efficiency, effectiveness, and computational cost. For instance, COTH (color thresholding) might be prioritized for colorful magazine covers, while ADTH (adaptive thresholding) is used for gray-level images. Conflicting evidence from different tools is reconciled using the Dempster-Shafer mathematical framework to produce a final hypothesis with a calculated confidence value.
Experimental Results
The system was evaluated using a custom database of 174 training images and 30 test images, specifically skewed toward difficult cases like hand-addressed mail and cluttered parcels. A benchmark "Mail Statistical Database" provided geometric features—such as typical aspect ratios and location probabilities—to guide the discrimination tools.
Performance Metrics
ABLS demonstrated robust performance across various mail classes. The overall success rate on the difficult training set was 81%, with the highest accuracy observed in standard letters (92%) and the lowest in irregular parcels (62%).
On a separate test set of 30 unseen, difficult images, the system maintained a 77% success rate. A critical insight from the results was that the primary bottleneck is initial image segmentation; when the system correctly segmented the address block, it succeeded in identification 98% of the time.
Discussion and Conclusion
While ABLS proved highly effective at locating addresses in complex environments, processing speed remains a significant hurdle for production environments. On a SUN-3 workstation, the average processing time was approximately 10.4 minutes per piece. However, the modular nature of the blackboard architecture makes the system an ideal candidate for parallelization using specialized hardware.
Future work focuses on improving robust "block splitting" tools to handle over-segmentation, where address blocks are accidentally merged with nearby text or graphics. By refining these initial segmentation routines, the overall system accuracy is expected to approach the near-perfect rates seen in correctly segmented cases.
Summary updated