渐进式特征增强的弱监督显著性目标检测

Progressively Feature-Enhanced Weakly Supervised for Salient Object Detection

针对多数弱监督显著性检测方法在复杂场景下容易出现目标结构缺损、边界粗糙等问题,提出一种渐进式特征增强的弱监督显著性检测算法。首先针对显著目标结构不完整问题,设计一种渐进式特征增强机制,主要包括双流语义增强模块和层次化自适应特征聚合模块,通过复用这种机制可以捕获更丰富的图像特征;其次为获取清晰完整的目标边缘,提出边缘引导模块,可以生成高质量的显著目标边缘图;最后将得到的边缘对显著区域预测网络进行指导,以生成结构完整且边界平滑的检测结果。在5个公开数据集上的实验结果表明,相比经典的WSSA算法,该算法在PASCAL-S数据集上平均绝对误差(MAE)降低了21.32%,F-measure值提高了6.27%,优于大多数先进的弱监督显著性目标检测算法。

For the weakly supervised object detection methods,it is easy to have problems such as object structure defect and rough boundary in complex scenarios.To overcome this problem,a progressively feature-enhanced,weakly supervised network for salient object detection has been proposed.First,this model uses a full convolutional neural network(ResNet-50)as the main stem for feature extraction,which helps the model learn the intrinsic features of the salient regions.To address the problem of the incomplete structure of salient objects,a Progressive Feature-Enhanced Mechanism(PFEM)is designed,which mainly includes a dual-stream semantic enhancement module and a hierarchical adaptive feature aggregation module.By reusing the progressive feature-enhancement mechanism,richer image features can be captured.Moreover,an Edge-Guided Module(EGM)is proposed to exploit the salient object edge structure.EGM can effectively aggregate the boundary formation of features and capture high-quality salient edge maps.Finally,salient edge maps are used to generate detection results with a complete structure and clear boundaries.Experimental results on five public datasets show that,compared with the classical WSSA algorithm,the Mean Absolute Error(MAE)on the PASCAL-S dataset is reduced by 21.32%,and the F-measure is increased by 6.27%,which is better than most advanced weakly supervised significance target detection algorithms.