复杂背景下的小目标检测算法

Small target detection algorithm in complex background

提出一种改进的多类别单阶检测器(SSD)算法.借鉴特征金字塔算法的思想,将Conv4-3层的特征与Conv7、Conv3-3层的特征进行融合,同时增加融合后特征图每个位置对应的默认框数量.在网络结构中增加裁剪-权重分配网络(SENet),对每层的特征通道进行权重分配,提升有用的特征权重并抑制无效的特征权重.为了增强网络的泛化能力,对训练数据集进行一系列增强处理.实验结果表明,改进后的算法在VOC数据集(07+12)上的检测效果良好,平均精度均值为80.4%,比改进前的算法提高了2.7%;在COCO数据集(2017)上的平均精度均值为42.5%,比改进前的算法提高了2.3%.所提算法能够准确检测出不小于16×16像素的目标.

An improved single-shot-multibox-detector(SSD)algorithm was proposed.Referring to the feature pyramid networks(FPN)algorithm,the features of the Conv4-3 layer were merged with the features of Conv7 and Conv3-3 layers,and the number of default boxes at each location in merged feature map was increased.The squeezeand-excitation networks(SENet)was added to the network structure;the feature channels of each layer were weighted,in order to enhance the useful feature weights and suppress the invalid feature weights.A series of enhancements were performed on the training data to enhance the generalization performance of the network.The experimental results show that the improved algorithm has a better performance on the VOC(07+12)dataset;the mean average precision(mAP)value of the improved algorithm is 80.4%,which is 2.7%higher than that of the original algorithm;the mAP value of the improved algorithm on COCO dataset(2017)is 42.5%,which is 2.3%higher than that of the original algorithm.Thus,the proposed algorithm can accurately detect the target with a size of at least 16×16 pixels.