基于空间-通道注意力的改进SSD目标检测算法

目标检测的任务是精确识别,有效定位出图像中目标物体,且预定义其类别.针对主流目标检测(single shot multibox detector,SSD)算法存在小目标检测准确度不高,检测效率较低等问题,提出一种基于空间-通道注意力机制的SSD目标检测算法(spatial and channel single shot multi-box detector,SC_SSD).通过在SSD深层网络引入空间-通道注意力机制增强高层特征图语义信息,提高算法获取目标物体的细节与位置信息的能力,从而降低漏检率及误检率,并提高小目标物体检测的准确度.此外,利用MobileNetV2中的深度可分离卷积对SSD主干网络(visual geom-etry group network,VGG-16)进行剪枝处理,降低参数量,从而减少训练与检测的时间.在PASCAL VOC 2007数据集上进行实验,本文算法检测的精确度与速度分别为78.9％与59.4 Fps,比SSD算法提升了 3.2％与26.7 Fps,满足实时性需求.算法也优于相比较的其他算法,是一种有效可行的目标检测算法.

Improved SSD object detection algorithm based on space-channel attention

The task of object detection is to identify accurately,effectively loc ate the target object in the image,and predefine its category.Aiming at the problems that the SSD (singal shot multibox detector) object detection algorithm is not enough in detection accurac y and efficiency in small object detection,an improved SSD algorithm based on space-channel att ention SC_SSD (spatial and channel singal shot multibox detector) is proposed.By introducing the space-channel attention mechanism into SSD deep network,the semantic informati on of high-level feature can be enhanced,and the algorithm is capable of capturing t he information of details and position of the target object.As a result,the rates of missed dete ction and false detection are reduced and the detection accuracy for small object is improved.I n addition,the deep separable convolution in MobileNetV2is used to prune the SSD backbone netw ork VGG-16(visual geometry group network).This will reduce the number of parameter s and will improve efficiency of training and detection.Experiments on Pascal VOC2007data set show that the detection precision and the detection rate of the proposed algorithm is 78.9% and 59.4Fps, respectively,3.2% and 26.7Fps higher than SSD algorithm,meeting the real-tim e requirements. The proposed algorithm is also superior to that of the previous ones and is an e ffective and feasible object detection algorithm.