基于深度学习的多视窗SSD目标检测方法

Object detection method of multi-view SSD based on deep learning

提出了一种基于深度学习的多视窗SSD(Single Shot multibox Detector)目标检测方法。首先阐述了经典SSD方法的模型与工作原理,并根据卷积感受野的概念和模型特征层与原始图像的映射关系,分析了各层级卷积感受野大小和特征层上默认框在原始图像上的映射区域尺寸,揭示了经典SSD方法在小目标检测上不足的原因。基于此,提出了一种多视窗SSD模型,阐述了其模型结构与工作原理,并通过106张小目标图像数据集测试,评估和对比了多视窗SSD方法与经典SSD方法在小目标检测上的物体检索能力与物体检测精度。结果表明:在置信度阈值为0.4的条件下,多视窗SSD方法的AF(Average F-measure)为0.729,m AP(mean Average Precision)为0.644,相比于经典SSD方法分别提高了0.169和0.131,验证了所提出算法的有效性。

The object detection method of multi-view Single Shot multibox Detector（SSD） based on deep learning was proposed. Firstly, the model and the working principle of classical SSD were expounded. According to the concept of convolution receptive field and the mapping relationship between the feature map and the original image, the sizes of covolution receptive field in different levels and the scales of the default boxes mapped to the original image were analyzed to find the reason why the classical SSD was not good at small object detection. Based on this, the multi-view SSD model was put forward, and the model architecture and its working principle were deeply expounded. Then, through the test in a dataset of 106 images for small object detection, the detection performance of multi-view SSD and classical SSD were evaluated and compared in object retrieval ability and object detection precision. Experimental results show that with the confidence threshold of 0.4, the multi-view SSD is 0.729 in Average F-measure（AF） and 0.644 in mean Average Precision（mAP）, and has respectively raised 0.169 and 0.131 compared to the classical SSD in the two evaluation indexes, thus verifying the effectiveness of the proposed method.