Approach to Improve Detection Model for Small Object in Complex Scenes

doi:10.3778/j.issn.1002-8331.2012-0521

Abstract

Abstract: Small object detection in complex scenes is a research difficulty and hot spot in the field of object detection. Many current one-stage and two-stage detection models both divide the positive and negative sample sets by pre-setting the intersection over union（IoU） threshold of the anchor box and the real object box. The anchor box is also used to obtain candidate boxes, and then to obtain the detection result. However, in complex scenarios, the preset IoU threshold leads to the imbalance of positive and negative samples. In addition, it is difficult to guarantee the location and density of the covered target with the predefined anchor boxes. To avoid the above problems, this paper proposes an adaptive anchor box（AAB） method to optimize the object detection network, and uses a clustering algorithm based on shape similarity distance to generate anchor boxes to improve object position accuracy. The anchor box of the class calculates the adaptive threshold selection（ATS）, and divides the positive and negative samples to ensure sample balance. The experimental results show that the detection of small targets（ship targets） in complex scenes, the object detection model of adaptive anchor frame method and adaptive threshold selection method can improve the accuracy of detection in complex scenes. Compared with faster R-CNN, FPN, Yolo3, and pp-Yolo, the models that incorporate the above new methods have improved the detection accuracy rate by 9.6, 2.6, 9.8 and 9.9?percentage points respectively.

Key words: small object detection, adaptive threshold selection, adaptive anchor boxes

摘要： 复杂场景下小目标检测是目标检测领域的研究难点和热点。传统的two-stage和one-stage检测模型都是通过预先设定锚点框与真实目标框的交并比（intersection over union，IoU）阈值来划分正负样本集，同时这组预定义的固定锚点框还用于获取候选框，进而得到检测结果。然而，在复杂场景下，预先设定的IoU阈值会带来正负样本不均衡问题；针对小尺寸目标（船舶）检测，预定义的锚点框也很难保证覆盖目标的位置和密度，因此限制了检测模型的准确率。为了解决上述问题，提出自适应锚点框（adaptive anchor boxes，AAB）的方法优化目标检测网络，采用基于形状相似度距离的聚类算法生成锚点框，提高目标区域定位技术；采用利用聚类的锚点框计算自适应IoU阈值（adaptive threshold selection，ATS），划分正负样本，保证样本均衡。对复杂场景下的小目标（船舶目标）进行检测，实验结果表明，采用自适应锚点框方法和自适应阈值选择方法的目标检测模型在复杂场景中检测均能提升准确，对比faster R-CNN、FPN、Yolo3和pp-Yolo，融合了上述新方法的模型均提升了检测准确率，分别提升了9.6、2.6、9.8和9.9个百分点。

关键词: 小目标检测, 自适应阈值选择, 自适应锚点框

ZHOU Hui, YANG Fenglong, CHU Na, LIU Zhenyu. Approach to Improve Detection Model for Small Object in Complex Scenes[J]. Computer Engineering and Applications, 2022, 58(11): 187-192.

周慧, 严凤龙, 褚娜, 刘振宇. 一种改进复杂场景下小目标检测模型的方法[J]. 计算机工程与应用, 2022, 58(11): 187-192.

References

[1] ZOU Z，SHI Z，GUO Y，et al.Object detection in 20 years：a survey[J].arXiv：1905.05055，2019.
[2] OKSUZ K，CAM B C，KALKAN S，et al.Imbalance problems in object detection：a review[J].IEEE Transactions on Pattern Analysis and Machine Intelligence，2021，43（10）：3388-3415.
[3] REN S，HE K，GIRSHICK R，et al.Faster R-CNN：towards real-time object detection with region proposal networks[J].IEEE Transactions on Pattern Analysis & Machine Intelligence，2017，39（6）：1137-1149.
[4] LIN T Y，DOLLáR P，GIRSHICK R，et al.Feature pyramid networks for object detection[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition，Hawaii，2017：2117-2125.
[5] REDMON J，FARHADI A.Yolov3：an incremental improvement[J].arXiv：1804.02767，2018.
[6] LONG X，DENG K，WANG G，et al.PP-YOLO：an effective and efficient implementation of object detector[J].arXiv：2007.12099，2020.
[7] LIU W，ANGUELOV D，ERHAN D，et al.SSD：single shot multibox detector[C]//Proceedings of the European Conference on Computer Vision，Amsterdam，the Netherlands，October 11-14，2016.Berlin，Heidelberg：Springer，2016：21-37.
[8] KANG M，LENG X，LIN Z，et al.A modified faster R-CNN based on CFAR algorithm for SAR ship detection[C]//Proceedings of the 2017 International Workshop on Remote Sensing with Intelligent Processing（RSIP），Shanghai，China，May 18-21，2017：1-4.
[9] XIAO X，ZHOU Z，WANG B，et al.Ship detection under complex backgrounds based on accurate rotated anchor boxes from paired semantic segmentation[J].Remote Sensing，2019，11（21）：2506.
[10] WANG Y，WANG C，ZHANG H.Combining a single shot multibox detector with transfer learning for ship detection using sentinel-1 SAR images[J].Remote Sensing Letters，2018，9：780-788.
[11] HU C H，CHEN C，HE C，et al.Ship small target detection based on deep convolution neural network in SAR image[J].Journal of Chinese Inertial Technology，2019，27（3）：397-405.
[12] ZHANG S，CHI C，YAO Y，et al.Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition，2020：9759-9768.
[13] HUANG Z，PAN Z，LEI B.Transfer learning with deep convolutional neural network for SAR target classification with limited labeled data[J].Remote Sensing，2017，9：907.
[14] KE W，ZHANG T，HUANG Z，et al.Multiple anchor learning for visual object detection[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition，2020：10206-10215.
[15] TIAN Z，ZHAN R，HU J，et al.Generating anchor boxes based on attention mechanism for object detection in remote sensing images[J].Remote Sensing，2020，12（15）：2416.
[16] GAO M，DU Y，YANG Y，et al.Adaptive anchor box mechanism to improve the accuracy in the object detection system[J].Multimedia Tools and Applications，2019，78（19）：27383-27402.
[17] YANG T，ZHANG X，LI Z，et al.MetaAnchor：learning to detect objects with customized anchors[C]//NeurIPS，2018：320-330.
[18] WANG J Q，CHEN K，YANG S，et al.Region proposal by guided anchoring[C]//IEEE CVPR，Long Beach Convention & Entertainment Center，Jun 15-21，2019：2965-2974.
[19] CHOI J，CHUN D，KIM H，et al.Gaussian yolov3：an accurate and fast object detector using localization uncertainty for autonomous driving[C]//IEEE ICCV，Seoul，Korea Oct 27-Nov 3，2019：502-511.
[20] 王光，林国宇.改进的自适应参数DBSCAN聚类算法[J].计算机工程与应用，2020，56（14）：45-51.
WANG G，LIN G Y.Improved adaptive parameter DBSCAN clustering algorithm[J].Computer Engineering and Applications，2020，56（14）：45-51.