结合多层次监督与边界损失的显著性目标检测

Saliency Object Detection Combining Multi-Level Supervision and Boundary Loss

针对PoolNet网络多次下采样操作易降低空间分辨率、仅使用二分类交叉熵损失函数(BCE)不利于捕捉显著性目标边缘特征,易导致边缘检测精度不高的问题。通过在PoolNet网络解码器的5级特征融合模块(fuse)之后分别加入多层次监督模型(MSM),并对5级MSM的输出分别按照BCE、SSIM与IoU三个损失函数计算值的和作为边界损失函数值,最后将5级边界损失函数值的平均值作为网络的最终输出损失值,并按照随机梯度下降法进行学习。从而提出一种结合多层次监督与边界损失的显著性目标检测方法:PoolNet-D。在6个常用数据集上的对比实验结果表明,提出的PoolNet-D模型在MAE和F-measure评价指标方面均有明显提升。

Multiple downsampling operations in PoolNet network can easily reduce spatial resolution,and using only binary cross entropy loss function(BCE)is not conducive to capturing salient target edge features,which can easily lead to low edge detection accuracy.In this paper,the multi-level supervision model(MSM)is added after the five-level feature fusion module(fuse)of the PoolNet network decoder,and the sum of the calculated values of BCE,SSIM and IoU loss functions is used as the boundary loss function value for the output of the five-level MSM.The av⁃erage value of the five-level boundary loss function value is used as the final output loss value of the network,and it is learned using the random gradient descent method.Thus,a new saliency object detection method combining multilevel supervision module with boundary loss function,namely,PoolNet-D,is proposed.The comparative experimental results show that on the six commonly used salient object detection datasets,the PoolNet-D model has significantly improved in terms ofMAE and F-measure evaluation indicators.