Key points are not available for this paper at this time.
The detection of Oracle Bone Inscription (OBI) is one of the most fundamental aspects of oracle bone morphology. However, the detection method depending on experts' experience requires longterm learning and accumulation for professional knowledge. This paper investigated the performance of the deep-learning-based object detection framework in the OBI dataset, then selected the one with the best performance as the baseline and made a series of optimization. Specifically, we first redesigned the sizes and ratios of the anchor box according to the data characteristics by using K- means clustering. Secondly, we extracted some typical noises from OBI for data augmentation. Finally, Focal Loss and Mixed-precision are used to improve the model precision and compress the memory footprint. To further improve the performance, the Spatial Pyramid Block is proposed, which can stabilize features and suppress noise interference. Experiments on our OBI benchmarks validate the superiority of the proposed method that achieves 82.1% F-measure suppressing several mainstream object detectors. Our dataset and algorithms will soon be available at http://jgw.aynu.edu.cn.
Liu et al. (2020) studied this question.