面向小目标检测结合特征金字塔网络的SSD改进模型

摘要

针对SSD卷积神经网络模型对小目标检测精度不高的问题,提出了一种基于特征金字塔网络的SSD改进模型.特征金字塔网络可以将深层的携带有更抽象、更丰富的语义信息的卷积特征图与浅层的分辨率更高、更细节的卷积特征图进行融合.检测的过程是将原始SSD网络得到的多层特征图,经改进设计的横向连接层、上采样层、融合层和预测层处理后,再通过非极大值抑制得到最终的检测结果.采用PASCAL VOC 2007和2012( train+val)作为训练集,PASCAL VOC 2007(test)测试集的mAP达到了75.8%,相比原SSD模型提高了1.5%.其中,在盆栽植物类密集小目标检测上有9.9%的提升.%To solve the problem that the low accuracy of SSD convolution neural network model for small target detection, an improved SSD model based on feature pyramid network was proposed. The feature pyramid network could fuse the deeper convolutional feature maps, which had more abstract and richer semantic information, and the shallower convolutional feature maps, with higher resolution and more de-tailed information. The detection process was that multi-layer feature maps obtained from the original SSD network were processed by the lateral connection layer, upsampling layer, fusion layer, and prediction layer and so on. And then the final detection results were achieved by the non-maximal suppression. In the test, PASCAL VOC 2007 and 2012 ( train+val) were used as training sets. The mAP in the PASCAL VOC 2007 (test) test set reached 75.8%, which was 1.5% higher than the original SSD model. In parti-cular, there was a 9.9% improvement in dense small-object detection of potted plants.