全局信息引导的多尺度显著物体检测模型

GLOBAL CONTEXT GUIDED MULTI-SCALE FEATURE NETWORK FOR SALIENT OBJECT DETECTION

目前方法不能有效提取图像的多尺度特征并完全利用全局信息,对此提出一种新的全局信息引导的多尺度显著物体检测模型。设计了多尺度特征提升模块,提升了各个侧输出特征的多尺度表征能力;利用空洞空间卷积池化金字塔模块提取图像的全局特征;将全局特征直接与各个侧边输出的多尺度特征相融合,利用全局特征引导侧边特征聚焦于目标区域中有用的中低层特征;采用由粗至细的方式得到最终的预测结果。该模型在四个广泛使用的数据集上进行了测试,并与8种近3年具有代表性的方法进行比较。实验结果表明,该模型对各种场景的显著物体检测具有较好的鲁棒性,可以均匀高亮显著目标的同时抑制背景噪声。

Current methods cannot represent multi-scale features effectively and utilize global information completely.This paper proposes a novel salient object detection approach using global context guide.We designed a multi-scale feature enhancement module to improve the capability of feature representation of multi-level side-output features.We utilized the atrous spatial pyramid pooling module to obtain the global features.The global features were directly fused with the multi-scale side-output features,which guided the side features to focus on the useful middle and low level features in the target area.The finally estimated results were obtained in a coarse-to-fine manner.We compared our approach with 8 state-of-the-art methods on four public benchmark datasets.Results demonstrate the robustness of the proposed approach on various scenarios,and it can evenly highlight salient targets while suppressing background noise.