A coarse-to-fine approach for industrial meter detection and its application

Abstract

This paper introduces a coarse-to-fine approach for industrial meter detection. This work has two key contributions. First, our method describes a two-level cascaded regressor to directly regress the industrial meter's parameter representation with normalizing target images to the same pose and scale, avoiding searching in multi-scale space with sliding windows. Second, after normalization, our method proposes a post verifier to largely decrease the false positive rate while keeping the true positive rate relatively high. Considering real-time performance, our method runs at 15 frames/s without multi-thread acceleration, which is essential for practical application. Evaluating with various on-site data, this approach achieves 97.5% hit rate while keeping the false positive rate below 1.35%. What's more, when applying this detection method to meter reading, the accuracy of digits reading achieves 95.5%, and the accuracy of pointer indicator detection achieves 95.6%, while the average error of estimated pointer indicator reading is limited to 6.6% normalized by measure range. Our coarse-to-fine approach shows promising prospect in practical applications.