基于多尺度特征解码网络的RGB显著性目标检测

RGB Salient Object Detection Based on Multi-scale Feature Decoding Network

近年来,基于深度学习的显著性目标检测(SOD)取得了很大进展。目前主流的基于深度学习方法的RGB显著性目标检测,忽略了编码器和解码器之间信息交换,以及不同层级编码器对最终预测图的贡献差异。本文设计了一种基于多尺度特征解码的RGB显著性目标检测网络,通过在编码和解码模块之间增加精炼过渡层和注意力机制,对编码器输出特征进行打磨,以还原更详细的显著性信息。此外,在网络的顶部增加感受野增强模块,以定位不同尺度信息,增强深层特征的全局语义信息,使预测结果更准确。在主流的6个数据集上的测试结果显示,本文的方法优于其他同类算法。

In recent years,salient object detection(SOD)based on deep learning has made great progress.The current mainstream RGB SOD based on deep learning methods always ignore the information exchange between encoder and decoder and the difference in the contribution of different layers of encoder to the final prediction map.In this paper,we design a multi-scale feature decoding network for RGB SOD to polish the encoder output features by adding a refining transition layer and an attention mechanism between the encoding and decoding modules to restore more detailed saliency information of each layer encoder.In addition,a receptive field enhancement module is added at the top of the network to locate different scale information and enhance the global semantic information of deep features leads to more accurate prediction results.The test results on 6 mainstream datasets show that the method in this paper outperforms the others.