Multiscale object detection based on channel and data enhancement at construction sites

作     者：Wang, Hengyou Song, Yanfei Huo, Lianzhi Chen, Linlin He, Qiang 

作者机构：Beijing Univ Civil Engn & Architecture Sch Sci Beijing 100044 Peoples R China Chinese Acad Sci Aerosp Informat Res Inst Beijing 100094 Peoples R China Beijing Univ Civil Engn & Architecture Inst Big Data Modeling & Technol Beijing 100044 Peoples R China 

出 版 物：《MULTIMEDIA SYSTEMS》 (多媒体系统)

年 卷 期：2023年第29卷第1期

页      面：49-58页

核心收录：

学科分类：08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：National Natural Science Foundation of China [62072024, 41971396, 61971290] Research Ability Enhancement Program for Young Teachers of Beijing University of Civil Engineering and Architecture [X21024] Outstanding Youth Program of Beijing University of Civil Engineering and Architecture BUCEA Post Graduate Innovation Project R&D Program of Beijing Municipal Education Commission [KM202110016001, KM202210016002] 

主　　题：Multiscale object detection Data enhancement Feature pyramid Subpixel convolution Channel enhancement 

摘      要：Object detection based on computer vision techniques plays an important role in the safety monitoring of large-scene construction sites. However, current object detection algorithms typically have poor performance on small targets. In this study, an enhanced multiscale object detection algorithm is developed to solve the problem of poor detection performance due to scale changes at construction sites. First, a scale-aware data automatic augmentation is defined to learn a data augmentation strategy. Then, to mitigate information loss caused by channel reduction when using feature pyramid network, we propose a method based on subpixel convolution to perform channel enhancement and upsampling, and add a bottom-up path to enhance the complete feature hierarchy with accurate localization signals in the lower layers. Experimental results show that the proposed algorithm achieves better accuracy on the construction site (MOCS) data set and the MS COCO data set. For example, compared with the Faster R-CNN detector with the ResNet-50 backbone network on the MOCS data set and MS COCO data set, the average accuracy increased by 8.0% and 1.5%, respectively. In particular, the average accuracy of small targets increased by 10.3% and 3.4%, respectively.