基于反卷积和特征融合的SSD小目标检测算法

SSD small target detection algorithm based on deconvolution and feature fusion

由于小目标的低分辨率和噪声等影响,大多数目标检测算法不能有效利用特征图中小目标的边缘信息和语义信息,导致其特征与背景难以区分,检测效果差。为解决SSD(single shot multibox detector)模型中小目标特征信息不足的缺陷,提出反卷积和特征融合的方法。先采用反卷积作用于浅层特征层,增大特征图分辨率,然后将SSD模型中卷积层conv112的特征图上采样,拼接得到新的特征层,最后将新的特征层与SSD模型中固有的4个尺度的特征层进行融合。通过将改进后的方法与VOC2007数据集和KITTI车辆检测数据集上的SSD和DSSD方法进行比较,结果表明:该方法降低了小目标的漏检率,并提升整体目标的平均检测准确率。

Given the low resolution and noise of small targets,most target detection algorithms cannot effectively utilize the edge and semantic information of small targets in feature maps,which makes it difficult to distinguish the features from the background.Thus,the detection effect is poor.To solve the problem of insufficient feature information of small and medium targets in the single shot MultiBox detector(SSD)model,we propose a method based on deconvolution and feature fusion.First,deconvolution is employed to process the shallow feature layer to increase the resolution of the feature graph.Then,the feature map of the convolution layer conv112 in the SSD model is sampled and spliced.Subsequently,a new layer of features is obtained.Finally,the new layer of features is combined with the feature layer of the four scales inherent in the SSD model.The improved method is compared with the SSD and DSSD methods on the VOC2007 dataset and KITTI vehicle detection dataset.The results show that the method reduced the missed detection rate of small targets and improved the average detection accuracy of all targets.