高校化学工程学报2023，Vol.37Issue(6)：962-970,9.DOI:10.3969/j.issn.1003-9015.2023.06.011
基于ASPP-SOLOv2的复杂场景下透明玻璃仪器实例分割
Instance segmentation of chemical transparent glassware in complex scene using ASPP-SOLOv2
摘要
Abstract
Deep learning methods are not good for the identification of laboratory transparent glassware in complex scenes.An instance segmentation dataset for laboratory complex scenes containing 1548 images of frequently used chemical transparent glassware was established in this study.An instance-level segmentation method of transparent glassware based on SOLOv2 was proposed,which combined multi-scale information with atrous spatial pyramid pooling(ASPP),and the low-level feature representation was improved by bottom-up enhancement.The final average precision with more than 50%intersection over union is 90.50%and class mean average precision(APav)is 76.00%,which is 8.7%higher than the average accuracy of the baseline method.Ablation experiments show that ASPP can enhance the representation of geometric,edge features of transparent glassware,and improve the instance segmentation accuracy with dense overlapping objects.The mAP of proposed method is increased by 22.58%,and adding ASPP in the C3 stage of the backbone network can achieve the best balance of low-level information and high-level semantic information.关键词
实例分割/透明玻璃仪器数据集/动态快速实例分割/空洞空间金字塔池化Key words
instance segmentation/chemical transparent glassware dataset/segmenting objects by locations/atrous spatial pyramid pooling分类
信息技术与安全科学引用本文复制引用
葛建统,杨鑫,祝模芮,冉进业,翟持,张浩..基于ASPP-SOLOv2的复杂场景下透明玻璃仪器实例分割[J].高校化学工程学报,2023,37(6):962-970,9.基金项目
国家自然科学基金(21806131) （21806131）
西南大学教育教学改革研究项目(2022JY009) （2022JY009）
重庆市研究生教育教学改革研究项目(yjg223123) （yjg223123）
