基于图像分类的弱监督RGBD图像显著性检测方法

Weakly-supervised RGBD Image Saliency Detection Based on Image Classification

彩色深度(Red Green Blue Depth, RGBD)图像不仅包含红绿蓝三通道的颜色信息,还包含深度信息,因此能提供更全面的空间结构信息.近年来,随着RGBD图像的广泛应用,基于RGBD的图像显著性检测方法相继被提出.为了更好地解决弱监督图像显著性检测方法中的跨模态数据融合问题,本文提出一种基于图像分类的弱监督RGBD图像显著性检测方法.首先,本文通过基于梯度的类别响应机制生成初始类别响应图,同时使用传统的显著图检测算法生成初始显著图.然后,根据本文提出的基于深度图的优化策略将初始类别响应图和初始显著图融合形成伪标签.最后,通过本文提出的由加权交叉熵损失、条件随机场推理损失以及边缘损失构成的混合损失对网络模型进行训练.实验表明,本文提出的弱监督RGBD图像显著性检测方法具有先进的性能.

RGBD image contains not only the color information of red, green and blue channels, but also the depth information, therefore it can provide more comprehensive spatial structure information.In recent years, with the widespread application of RGBD images, someimage saliency detection methods based on RGBD images have been proposed.In order to solve the cross-modal data fusion problem in the weakly-supervised image saliency detection method, this paper proposes a weakly-supervised RGBD image saliency detection method based on image classification.First of all, this paper uses the gradient-based category response mechanism to generate the initial category response map, and at the same time uses the traditional saliency map detection algorithm to generate the initial saliency map.Then, according to the depth mapbased optimization strategyproposed in this paper, the initial category response map and the initial saliency map are fused to form a pseudo-label.Finally, the network model is trained through the mixed loss composed of weighted cross-entropy loss, conditional random field inference loss and edge loss proposed in this paper.Experiments show that the weakly-supervised RGBD image saliency detection method proposed in this paper has advanced performance.