基于多尺度信息处理和Gabor初始化的图像显著性对象检测

基于全卷积网络的图像显著性检测获得了广泛的关 注,并取得了令人瞩目的检测性能 。然而,该类型神经网络依然存在许多问题,如高复杂网络导致难以训练、显著性对象边缘 结果不准确等。针对这些问题,本文提出基于Gabor初始化的卷积神经网络。该网络主要特 点包括:1) 利用Gabor特征初始化卷积神经网络,提高神经网络训练效率； 2) 构建多尺 度 桥接模块,有效衔接编码和解码阶段,进而提高显著性检测结果； 3) 提出加权交叉熵损失 函数,提高训练效果。实验结果表明,本文提出的神经网络在三个不同的数据集上均显示出 优异的显著性对象检测性能。

Image salient object detection based on multi-scale information processing and Gabor initialization

FCN based image saliency detection met hods have become very popular in the past few years because of outstanding performance.However,problems still exist in the methods,such as difficult training process,inaccurate edge detect ion.To solve the problems,a convolutional neural network based on Gabor initia lization is proposed,and the features of the network include:1) making use of Gabor feature improving the efficiency of training process; 2) constructing brid ge module for the purpose connecting coding and decoding processes; 3) proposing weighted cross entropy loss function to facilitate training effect.The experim ent shows that the proposed network represent brilliant performance in three pub lic datasets.