基于PDSSD改进型神经网络的小目标检测算法

SMALL OBJECT DETECTION ALGORITHM BASED ON PDSSD IMPROVED NEUAL NETWORK

SSD卷积神经网络一直对较小目标检测精度不佳。对此在SSD网络结构的基础上引入空洞卷积(Dilated Convolution),并组建空洞金字塔模块(Pyramid Dilated Convolution)和特征空洞金字塔模块(Feature Pyramid Dilated Convolution)融入SSD中,提升了网络浅层特征层的语义信息,提高了深层特征层的感受野和特征提取能力,构建了新型网络结构Pyramid Dilated SSD(PDSSD)。实验结果表明,PDSSD在PASCAL-VOC数据集上的检测mAP(Mean Average Precision)值高达82.1%,检测精度和小目标检测能力明显高于SSD,并且网络训练速度和mAP值领先于其他主流算法。

The SSD convolutional neural network is used to detect small objects with poor accuracy.Based on the SSD network structure,we introduce the dilated convolution,and construct the pyramid dilated convolution and feature pyramid dilated convolution into SSD.It improves the semantic information of the shallow feature layer,improves the receptive field and feature extraction ability of the deep feature layer,and constructs a new network structure,pyramid digested SSD(PDSSD).The experimental results show that the mean average precision(mAP)of PDSSD on PASCAL-VOC dataset is as high as 82.1%,the detection accuracy and small target detection ability of PDSSD are significantly higher than SSD,and the network training speed and mAP value are ahead of other mainstream algorithms.