关键词: 目标检测, 转置卷积, 特征聚合, 单阶段多边框目标检测模型

Abstract: Since the mean Average Precision (mAP) of Single Shot multibox Detector (SSD) drops significantly when evaluating with higher Intersection over Union (IoU), a feature aggregation method using transposed convolution as main component was proposed. On the basis of SSD model, a deep Residual convolutional Network (ResNet) with 101 layers was used to extract features. Firstly, abstraction of semantics and context information was generated by using transposed convolutional layers which doubled the scales of deeper feature maps. Secondly, fully connected convolutional layers were applied to shallow layers to prevent unexpected bias. Finally, the shallow and deep feature maps were concatenated together, and convolutional layers with kernel size 1 were used to reduce the channel sizes. The feature aggregation can repeat multiple times. The experiments were conducted on KITTI dataset and took 0.7 as IoU threshold. Experimental results show that the mAP was improved by about 5.1 and 2 percent points compared to the original SSD model and the state-of-the-art Faster R-CNN model. The feature aggregation model can effectively improve the mAP and generate high quality bounding boxes in object detection tasks.

Key words: object detection, transposed convolution, feature aggregation, Single Shot multibox Detector (SSD) model