YOLO-Anti: YOLO-based counterattack model for unseen congested object detection. (November 2022)

• Record Type: Journal Article
• Authors: Wang, Kun; Liu, Maozhen
• Is Part Of: Pattern recognition. Volume 131(2022)
• Publication Date: 2022-11
• Subjects: Deep learning -- Congested and occluded objects -- Object detection; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.108814
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: The paper proposes a novel and efficient network framework dedicated to unseen congestion detection. An adaptive context module similar to valve control is proposed to obtain contextual information that balances foreground and background features. The paper demonstrates that balanced features on multi-level network prediction maps have a prevailing impact on the accuracy of object detection. Accurate detection results are realized through the proposed anti-congestion model for flexible regional mapping. Abstract: Object detection is advancing rapidly with the development of deep learning solutions and big data dimensions. This paper takes the challenging recognition task as the core work and proposes a novel and efficient network framework dedicated to unseen congestion detection. To guarantee the accuracy as well as the speed of inference, the detector utilizes the advanced You Only Look Once v4 (YOLOv4) as the backbone and agglutinates the four proposed strategies, called YOLO-Anti. Our model mainly consists of three modules: First, an adaptive context module similar to valve control is proposed to obtain contextual information that balances foreground and background features. Second, to solve the problem that the imbalance between feature levels weakens the detection performance, a balanced prediction layer method is developed. Finally, we propose an anti-congestion network to selectively expand the local domain to achieve finer-grained detection. Besides, in the training procedure, a designed heterogeneous cross-entropy loss is utilized to strengthen the detector's discrimination of similar targets in different categories. Extensive experiments were conducted on the PASCAL VOC, COCO, and UA-DETRAC data sets. The state-of-the-art results were achieved on UA-DETRAC and the leading performance on PASCAL VOC and COCO. Also, compared with baseline YOLOv4, the proposed method brings significant accuracy improvement and negligible time consumption.