Improved single shot detection using DenseNet for tiny target detection

• DOI: 10.1002/cpe.7491
• Publication Date: 2022-11-07T01:56:59Z
• Authors: Shudi Wang; Manman Xu; Ying Sun; Guozhang Jiang; Yaoqing Weng; Xin Liu; Guojun Zhao; Hanwen Fan; Jun Li; Cejing Zou; Yuanmin Xie; Li Huang; Baojia Chen
• Keywords: Pascal (unit); Backbone network; Feature (linguistics); Single shot; Convolution (computer science)

Summary As the development of deep learning and continuous improvement computing power, as well needs social production, target detection has become a research hotspot in recent years. However, algorithm problem that it is more sensitive to large targets does not consider feature‐feature interrelationship, which leads high false or missed rate small targets. An method (C‐SSD) based on improved SSD proposed, replaces backbone network VGG‐16 with dense convolution (C‐DenseNet) achieves further feature fusion through fast connections between blocks. The Introduction residuals prediction layer DIoU‐NMS improves accuracy. Experimental results demonstrate C‐SSD outperforms other networks at three different image scales best performance 83. A 8% accuracy PASCAL VOC2007 test set, proving effectiveness algorithm. better balance speed accuracy, showing excellent rapid