基于多尺度交互式残差结构的图像显著目标区域检测

Sailent object detection based on multi-scale interactive residual structure

为了解决显著目标检测中不同深度特征信息的有效提取与融合问题,设计了一种基于多尺度交互式残差模块的图像显著目标区域检测模型M-IRNet。基于深度学习算法框架,搭建了多尺度交互式残差结构模块IRM,提取具有高表征能力的多尺度特征信息,抑制噪声干扰,获取表征细节信息的低层特征信息和深层语音信息;辅助以双方向传播策略,并将不同深度的上下文信息有效融合,浅层的细节信息融入较深层信息,同时高层的语义信息调节低层信息,实现图像显著目标的定位与检测。在公开数据集上的实验结果表明,所设计的图像显著目标区域检测模型具有一定的优势,在PASCAL-S数据集上MAE降到了0.092,而在DUT-OMROM数据集上F-measure升到了0.763。通过提取更有效的多尺度特征,增加浅层和深层的特征比率,不仅提高了显著目标的细节特征表示能力,同时还增加了深层语义信息的定位能力,使应用于图像显著目标检测模型检测精度提升。

In order to effectively extract and integrate multi-scale information with different depths in salient object detection,the salient object region detection model,M-IRNet,based on multi-scale interactive residual module was designed.A multi-scale interactive residual structure module,IRM,was built to extract multi-scale feature information,to suppress noise interference,to obtain detail information in low-level feature and deep context information.For effective information fusion,a modified bi-directional propagation strategy was adopted,which can effectively fuse context information of different depths,and at the same time,the semantic information of the upper level regulates the information of the lower level to realize the location and detection of the significant object of the image.The experimental results on five benchmark datasets show that the designed model of image salient target region detection has certain advantages,especially on the PASCAL-S datasets the MAE drops to 0.092,and on the DUT-OMROM datasets,the F-measure drops to 0.763.According to extracting more effective multi-scale features to increase the ratio of shallow and deep features,not only the ability to represent the detailed features for salient object is improved,but also the ability to locate salient object is increased.It is applied to the image salient target detection model,which can improve the detection accuracy.