Automatic watermeter digit recognition on mobile devices

Abstract

Automatic watermeter digit recognition in the wild is a challenging task, which is an application of scene text recognition in the field of computer vision. In this paper, we propose an automatic watermeter digit recognition approach on mobile devices which consists of digit detection and recognition. Specifically, we adopt Adaboost with aggregated channel features (ACF) to detect watermeter digital regions, where the computation is accelerated by the fast feature pyramid technology. Then a small attention bidirectional long short-term memory (BLSTM) is designed for end-to-end digit sequence recognition. Convolutional Neural network (CNN) is exploited to extract discriminative feature and BLSTM is able to capture the rich context in both directions within sequence data. Moreover, an attention mechanism is added to weight the most important part of incoming image features. We validate the performace of our approach on the collected complex dataset. It contains various watermeter images in real scenario which has illumination changes, messy environment, half-digit and blurring. It is observed that the proposed algorithm outperforms existing methods. Our approach runs 10 fps with 96.1% accuracy on HUAWEI Mate 8.