Abstract: Aiming at the problems that most existing RGB-D saliency detection methods are difficult to effectively extract multi-scale features in the feature extraction process and that the insufficient interaction of modal feature information in the decoding process leads to information loss, an RGB-D saliency detection method with feature enhancement and progressive decoding is proposed. First, to address the difficulty of multi-scale feature extraction, which leads to the low detection accuracy of the algorithm for multi-target and small-target objects, a multi-scale self-attentive feature enhancement module is constructed; second, in order to fully utilize the advantages of the two modal features, a cross-modal feature fusion module is constructed to enhance the information interaction between the different modal branches; and in order to solve the problem of information loss in the decoding process, a progressive weighted fusion decoder is proposed to fully integrate the feature information of the different modalities and to refine the feature information. information of different modalities; finally, initial salient map optimization is achieved by supervised training of the layer-by-layer decoding results using a hybrid loss function. Through the comparison experiments with 10 advanced mainstream methods on five public benchmark datasets, it is found that the method performs better on all five metrics. The experimental results show that the method achieves a more advanced performance by obtaining relatively accurate significant maps in a wide range of complex situations.

Key words: image processing, saliency detection, deep learning, feature enhancement, progressive decoding

摘要： 针对现有的大多数RGB-D显著性检测方法在特征提取过程中难以有效提取多尺度特征以及在解码过程中模态特征信息交互不充分导致信息丢失的问题，提出了一种特征增强和渐进式解码的RGB-D显著性检测方法。首先，针对多尺度特征提取困难，导致算法对多目标、小目标物体检测精度低的问题，构建了多尺度自注意力特征增强模块；其次，为了充分发挥两种模态特征的优势，构建了跨模态特征融合模块，以增强不同模态分支之间的信息交互；为了解决解码过程中信息易丢失的问题，提出了渐进式加权融合解码器，充分整合细化不同模态的特征信息；最后，通过对逐层解码的结果采用混合损失函数监督训练，实现初始显著图优化。通过在5个公开基准数据集上与10种先进主流方法进行的对比实验表明，该方法在5个指标上均有较优表现。实验结果表明，该方法在多种复杂情况下都能获得相对精确的显著图，达到了较为先进的性能。

关键词: 图像处理, 显著性检测, 深度学习, 特征增强, 渐进式解码