1 Citation
This dzi beads recognition dataset collects original dzi beads pattern images with the help of both web crawling and manual mobile phone shooting. Among them, manual shooting uses smartphones to shoot and collect dzi beads images in different backgrounds, angles, sizes, and light and darkness levels; network crawling uses the Scrapy crawling framework to acquire data in Baidu Gallery, Google Gallery, Bing Gallery, and dzi collection website. In order to improve the quality, validity and category balance of the collected data, we consulted and referred to the classic works in the field of dzi beads, referred to the specifications of datasets such as ImageNet for image classification of deep learning, and combined with the experts in the field of dzi bead images collected for category screening and manual annotation, to form the Bagua dzi beads (baguatz), Baobao dzi beads (baopingtz), Caishen dzi beads (caishentz), Fenghuang dzi beads (caihentz) and Dzi bead images. Caishentz), fengyantz, guanyintz, guijiatz, huyatz, jinxiangoutz, jingangchutz, guirentz, and lianhuafaatz. Dzi (lianhuafaqitz), Ping An Cross Dzi (pinganshizitz), Bodhi Dzi (putitz), Longyantz Dzi (longyantz), Lotus Dzi (lianhuatz), Wanzitz Dzi (wanzitz), Ruyitz Dzi (ruyitz), Lightning Five-Eyed Dzi ( shandianwuyantz), water pattern dzi (shuiwentz), heaven and earth dzi (tianditz), sun, moon, and star dzi (ryuexingtz), and medicine dzi (yaoshitz), and other 5,443 original dzi beads pattern datasets in 22 categories gZiBeads . In addition, in order to improve the robustness and generalisation ability of the constructed deep learning dzi pattern automatic recognition model, the original dzi pattern dataset is subjected to data enhancement operations such as random cropping, flipping, panning, scaling and Gaussian blurring, and finally 8743 dzi pattern datasets of 22 dzi pattern categories are formed. In order to further evaluate the quality of the constructed dzi beads pattern dataset, the original dzi beads pattern dataset was trained and validated on mainstream deep learning image classification models such as InceptionV3, ResNet50, MobileNetv2, EfficientNetV2, and MobileViT, and the validation accuracy rate of MobileViT reaches 89.18%, meanwhile, the accuracy rate is as high as 93.8% when validated again on the data-augmented dataset.The experimental results show that data augmentation can help to improve the robustness and generalisation performance of the automatic dzi pattern recognition,...