A comprehensive comparison of end-to-end approaches for handwritten digit string recognition

Over the last decades, most approaches proposed for handwritten digit string recognition (HDSR) have resorted to segmentation, which is dominated by heuristics, thereby imposing substantial constraints on final performance. Few of them been based segmentation-free strategies where each pixel column has a potential cut location. Recently, added another perspective problem, leading promising results. However, these still show some limitations when dealing with large number touching digits. To bridge resulting gap, in this paper, we hypothesize that digits can be approached as sequence objects. We thus evaluate different end-to-end solve HDSR particularly two verticals: those object-detection (e.g., Yolo and RetinaNet) sequence-to-sequence representation (CRNN). The main contribution work lies its provision comprehensive comparison critical analysis above mentioned five benchmarks commonly used assess HDSR, including challenging Touching Pair dataset, NIST SD19, real-world datasets (CAR CVL) ICFHR 2014 competition HDSR. Our results model compares favorably against models advantage having shorter pipeline minimizes presence heuristics-based models. It achieved 97%, 96%, 84% rate NIST-SD19, CAR, CVL datasets, respectively.