分段反卷积改进SSD的目标检测算法

Improved SSD Object Detection Algorithm Based on Segmented Deconvolution

针对当前SSD算法低层特征图语义信息不足导致存在小目标漏检以及误检的问题,提出一种基于分段反卷积改进SSD的目标检测算法SD-SSD(Segmented Deconvolution-Single Shot M ulti Box Detector).根据SSD模型低层特征图语义信息提取不足,高层特征图边缘信息丢失过多,本文重新设计了融合结构,不仅降低了计算过程中的参数数量,而且丰富了各个特征图的细节信息和语义信息;由于特征图反卷积的次数过多会增加噪声信息,本文将模型中高层特征图分成三段做分段反卷积与低层特征层融合;为增强小目标在模型下的检测效果,增加更低层次的特征图进行特征融合,着重检测小目标.在Pascal VOC2007测试集上进行验证,本文SD-SSD模型大幅度提高了小目标类别的AP值,mAP相对SSD模型和DSSD模型分别提高了4.30%和3.0%,相比目前主流单阶段目标检测算法,本文算法保持了较高的检测精度和检测速度.

The lack of semantic information in the low-level feature maps of current SSD algorithm leads to the problem of missed detection and false detection of small objects.An SD-SSD(Segmented Deconvolutional-Single Shot MultiBox Detector)object detection algorithm based on segmented deconvolution is proposed.According to the low-level feature map of the SSD model,the semantic information extraction is insufficient.There is too much loss of detailed information in high-level feature maps.This article redesigned the fusion structure.Not only reduces the number of parameters in the calculation process,but also enriches the detailed information and semantic information of each feature map.Too many times of feature map deconvolution will increase noise information.In this paper,the high-level feature map in the model is divided into three segments for segmented deconvolution and low-level feature layer fusion.To enhance the detection effect of small objects under the model.Add lower-level feature maps for feature fusion to strengthen the detection of small objects.The algorithm is verified on the Pascal VOC2007 test set.In this paper,the SD-SSD model significantly improves the AP value of the small objects.mAP increased by 4.30%and 3.0%respectively compared to SSD and DSSD models.Compared with the current mainstream single-stage object detection algorithm,the algorithm in this paper still has higher detection accuracy and detection speed.