摘要
为探讨数据集标注对Faster R-CNN识别连衣裙衣领的影响,以框选范围依次扩大的3种标准(即衣领;唇部至肩部;唇部至腰部,后二者中的唇部对应平面服装图中的衣领上端)框选连衣裙图像中的衣领位置,制成3个数据集。以Faster R-CNN算法训练3个数据集,其中预训练网络均选用AlexNet。结果表明:召回率与平均精度均值随框选范围的扩大均有提高,分别提高了14%和13%,较小的立领、圆领、方领和V领提高最明显,而整体准确率随框选范围扩大保持不变。为进一步提升识别效果,以GoogLeNet为预训练网络,训练框选范围最大的数据集,得到的召回率、准确率和平均精度均值进一步增至86%、81%和83%。Faster R-CNN识别连衣裙衣领具有较好的可行性,扩大数据集框选区域可提升Faster R-CNN自动识别较小领型的效果。
To explore the impact of dataset annotations on the recognition of dress collars by Faster RCNN,three datasets were made following three standards to label collar positions in dress images.The regions be labeled of these three standards were expanded in turn(i.e.,the collar itself,from the lip to the shoulder,from the lip to the waist.In the last two,the lip corresponds to the upper end of the collar in flat state clothing images).Faster R-CNN algorithm was used to train the three datasets respectively,and AlexNet was chosen as pre-training networks.The results showed that the recall rate and the mean average precision were improved with the enlargement of regions of interest by 14%and 13%respectively.The improvement of recognition of the smaller collars like the stand-up collar,the round collar,the square collar,and the V collar was the most obvious,while the overall accuracy remained unchanged with the enlargement of regions of interest.To further enhance the recognition results,GoogLeNet was used as the pre-training network and the dataset with the largest region in the three standards was selected in the training process.The recall rate,accuracy rate and mean average precision were further increased to 86%,81%,and 83%.Faster R-CNN has good feasibility in dress collars′identification,and expanding the region of interest in the dataset properly can improve the performance of automatic recognition of smaller collars by Faster R-CNN.
作者
张艳清
刘成霞
ZHANG Yanqing;LIU Chengxia(School of Fashion Design and Engineering,Zhejiang Sci-Tech University,Hangzhou 310018,China;Zhejiang Province Engineering Laboratory of Clothing Digital Technology,Zhejiang Sci-Tech University,Hangzhou 310018,China)
出处
《浙江理工大学学报（自然科学版）》
2021年第6期751-757,共7页
Journal of Zhejiang Sci-Tech University(Natural Sciences)
基金
浙江省自然科学基金(LY20E050017)。
作者简介
张艳清(1996-),女,湖南衡阳人,硕士研究生,主要从事数字化服装技术方面的研究;通信作者:刘成霞,E-mail:glorior_liu@163.com。
