Tiny object detection with context enhancement and feature purification. (January 2023)

• Record Type: Journal Article
• Authors: Xiao, Jinsheng; Guo, Haowen; Zhou, Jian; Zhao, Tao; Yu, Qiuze; Chen, Yunhua; Wang, Zhongyuan
• Is Part Of: Expert systems with applications. Volume 211(2023)
• Publication Date: 2023-01
• Subjects: Tiny-object detection -- Context enhancement -- Feature purification -- Dilated convolution; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.118665
• Languages: English
• ISSNs: 0957-4174

Abstract

Abstract: Tiny object detection is one of the challenges in the field of object detection, which can be applied in a variety of fields. Thanks to the advances in deep learning, significant improvement has been made in image object detection. However, the performance of tiny object detection still needs to be considerably enhanced. In this paper, we proposed a novel feature pyramid composite neural network structure comprising two modules: the context enhancement module (CEM) and feature purification module (FPM). The top-to-bottom input of the feature pyramid network into the multi-scale dilated convolution features in the CEM can augment the context information. When fusing multi-scale features in the FPM, the feature purification procedures for channel and space dimensions are employed to eliminate conflicting information, and tiny objects are more noticeable in contradictory information. Furthermore, a new data-enhancement strategy is introduced to increase the contribution of tiny objects in the loss function, which is named copy-reduce-paste and improves the balance of training samples. Overall, the experiments on the VOC dataset show that the AP s score can reach 16.9% and the IOU is 0.5:0.95 for the suggested method. The AP s score is more than 3.9% that of YOLOV4, more than 7.7% that of CenterNet, and more than 5.3% that of RefineDet. On TinyPerson dataset, our AP s score is more than 0.8% that of YOLOV5, providing a new alternative solution for the tiny-object detection research community. Highlights: Tiny target data enhancement is used to improve the contribution of its loss in training. Context enhancement is used to supplement the context information for tiny targets. Feature purification is used to filter the conflict information after feature fusion.