面向显著性目标检测的SSD改进模型

传统显著性目标检测方法常假设只有单个显著性目标，其效果依赖显著性阈值的选取，并不符合实际应用需求。近来利用目标检测方法得到显著性目标检测框成为一种新的解决思路。SSD模型可同时精确检测多个不同尺度的目标对象，但小尺寸目标检测精度不佳。为此，该文引入去卷积模块与注意力残差模块，构建了面向多显著性目标检测的DAR-SSD模型。实验结果表明，DAR-SSD检测精度显著高于SOD模型；相比原始SSD模型，在小尺度和多显著性目标情形下性能提升明显；相比MDF和DCL等深度学习框架下的方法，也体现了复杂背景情形下的良好检测性能。

An Improved SSD Model for Saliency Object Detection

Traditional saliency object detection methods, assuming that there is only one salient object, is not conductive to practical application. Their effects are dependent on saliency threshold. Object detection model provides a kind of new solutions. SSD can accurately detect multi-objects with different scales simultaneously, except for small objects. To overcome this drawback, this paper presents a new multi- saliency objects detection model, DAR-SSD, appending a deconvolution module embedded with an attention residual module. Experiments show that DAR-SSD achieves a higher detection accuracy than SOD. Also, it improves detection performance for multi- saliency objects on small scales, compared with original SSD, and it has an advantage over complicated background, compared with MDF and DCL, which also are deep model based methods.