改进的 SSD-ResNet 算法

单次多边界框检测器（single shot multibox detector, SSD）算法因其性能优良已被应用于许多场景中，但该算法对小目标物体的检测精度偏低，主要原因是高层的语义信息没有被充分利用。为解决该问题，文章将其基础网络替换为残差网络（residual network, ResNet），同时融合深浅层的特征信息来增强浅层特征图的语义信息，此外还引入注意力模块，保留更多的目标特征信息，抑制无关信息，进一步提升对小目标物体的检测效果。在 PASCAL VOC2007 数据集上进行实验测试，平均精度均值为 80.2%，优于其他 SSD 改进算法。由于增加了特征融合和注意力模块，所提算法检测速度有所下降，但相比于 SSD 改进算法，检测速度仍有明显的优势。

As an algorithm with better detection accuracy and speed, single shot multibox detector (SSD) has made great progress in many aspects. However, it cannot achieve a good detection effect for small objects because it does not make full use of high-level semantic information. Aiming at this problem, VGG is replaced with residual network (ResNet) as the backbone network, the feature fusion method is used to enhance the semantic information of the shallow feature map, at the same time, the attention module is introduced, which retains more object feature information and suppresses irrelevant information. Through the above methods, the detection effect of small objects can be enhanced. By experimenting on the PASCAL VOC2007 datasets, the validity of the proposed algorithm is proven, the mean average precision value of the algorithm is 80.2%, which is better than those of other improved SSD algorithms. Although the addition of feature fusion and attention modules to the algorithm can cause a decrease in the detection speed, it is still better than deconvolutional single shot detector (DSSD) algorithm.