基于特征残差融合的显著性检测网络

NETWORK FOR SALIENCY DETECTION BASED ON RESIDUAL FUSION OF FEATURES

当前的显著性检测任务得益于卷积神经网络模型的监督训练能够达到很好的效果,但是模型中的显著性特征如何有效地利用仍是一个关键的问题。不同层级的显著性特征信息融合能够达到互补的效果进而促进最终预测的效果,因此提出一个基于局部信息残差融合的网络架构。该结构是对局部范围的卷积层的特征进行残差式的融合,以此降低由于采样操作导致引入噪点的风险。再将融合的新特征图由深层递进式地传递到浅层并输出,进而获得最终的预测结果。

Benefitting from convolution neural network with supervised training,recent works of saliency detection achieves good results.However,it is still a core issue that how to effectively use the salient features in the model.We believe that the fusion of different levels of saliency feature information can complement each other and promote effect of the final prediction.In this paper,a network framework based on local information residual fusion is proposed.This framework was to fuse the features of the local convolution layer in the form of residual error,so as to avoid the risk of introducing noise due to too many sampling operations.The fused new feature map was transmitted from deep layer to shallow layer progressively,and the final prediction result was obtained.