基于改进SSD的少样本目标检测  

Few-shot Object Detection Based on Improved SSD

作　　者：毕忠勤 单美静[2] 刘志斌 徐富强 BI Zhong-qin;SHAN Mei-jing;LIU Zhi-bin;XU Fu-qiang(School of Computer Science and Technology,Shanghai University of Electric Power,Shanghai 200090,China;Department of Information Science and Technology,East China University of Political Science and Law,Shanghai 201620,China)

机构地区：[1]上海电力大学计算机科学与技术学院,上海200090 [2]华东政法大学信息科学与技术系,上海201620

出　　处：《计算机技术与发展》2023年第11期35-40,共6页Computer Technology and Development

基　　金：上海市地方能力建设项目(23010501500)。

摘　　要：目标检测作为深度学习的热点问题之一,在自动驾驶、行人识别、智能医疗、机器人视觉等多个领域有着广泛的应用前景。但现有的大部分目标检测模型都依赖于大规模的标注数据集来训练模型以保证目标检测的准确率,而在许多实际的应用场景中,大量数据的标注不仅耗费人力物力,而且需要大量专业人士的参与,在一定程度上限制了目标检测模型的实际应用。针对少样本目标检测的特殊要求,基于SSD网络提出了一种改进的少样本目标检测模型,提高了目标检测应用的适用性。首先,在SSD(Single Shot multiBox Detector)网络的基础上,用ResNet-50代替VGG作为特征网络,从而提升模型的特征提取能力。其次,通过引入残差单元避免了网络退化问题。最后,为了充分融合各层之间的语义信息和位置信息,用FPN(Feature Pyramid Networks)替换了原模型中间的两个特征层。基于改进SSD网络的目标检测模型在少样本数据集的检测结果中,mAP值达到了79.8%,比原始模型提高了2.6百分点。As one of the hot topics of deep learning,object detection has a wide application prospect in many fields,such as automatic driving,pedestrian recognition,intelligent medical treatment,and robot vision and so on.However,most of the existing object detection models rely on large-scale annotation data sets for model training to ensure the accuracy of target detection.In many practical application scenarios,the annotation of a large number of data not only consumes human and material resources,but also requires the participation of a large number of professionals,which limits the practical application of the object detection model to a certain extent.Aiming at the special requirements of few-shot object detection,we propose an improved few-shot object detection model based on SSD(Single Shot multiBox Detector)network,which improves the applicability of object detection applications.Firstly,the ResNet-50 is used instead of VGG as the feature network on the basis of SSD network to improve the feature extraction capability of the model.Secondly,the problem of network degradation is avoided by introducing residual element.Finally,in order to fully integrate the semantic information and location information between the layers,FPN is used to replace the two feature layers in the middle of the original model.In the detection results of the target detection model based on the improved SSD network in a few-shot data set,the mAP value of the improved model reached 79.8%,which was 2.6 percentage points higher than that of the original model.

关 键 词：目标检测 机器学习 少样本 FPN SSD网络模型 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]