光学学报, 2019, 39 (9): 0915005, 网络出版: 2019-09-09
双注意力循环卷积显著性目标检测算法 下载: 1522次
Salient Object Detection Algorithm Based on Dual-Attention Recurrent Convolution
机器视觉 显著性目标检测 全卷积神经网络 注意力机制 循环卷积网络 machine vision salient object detection fully convolutional neural network attention mechanism recurrent convolutional network
摘要
显著性目标检测是机器视觉领域的研究热点,具有广泛的应用前景。针对现有显著性目标检测算法存在的显著区域检测不均匀、边缘表示模糊等问题,提出一种双注意力循环卷积显著性目标检测算法。在U-Net全卷积骨干网络中添加像素间-通道间双注意力模块,在跨层连接前对底层特征进行预处理,减小噪声和杂波干扰,提高显著区域检测性能。在骨干网络后端使用循环卷积模块,将最后的预测图与底层卷积层特征进一步结合,增强预测区域边缘的表示效果。在三个公开数据集上进行实验评测,并与相关算法进行对比,结果表明所提算法能更好地均匀突显显著区域和细化区域边缘。
Abstract
Salient object detection has attracted considerable attention in the field of the machine vision, with a wide range of applications. This study proposes a salient object detection algorithm based on the dual-attention recurrent convolution to overcome the limitations associated with the existing algorithms, i.e., uneven salient region detection and fuzzy edge representations. A dual-attention module consisting of pixel- and channel-wise attentions is added to a backbone U-Net fully convolutional network to preprocess the shallow convolutional features before skip-layer connection, and reduce noise and clutter interference. This improves its salient region detection performance. Then, following the backbone network, a recurrent convolutional module enhances the edge representation of the prediction region by combining the final prediction map with the shallow convolutional features. The results of experiments on three open datasets show that the proposed algorithm is better able to highlight salient regions and refine their edges than other correlation algorithms.
谢学立, 李传祥, 杨小冈, 席建祥. 双注意力循环卷积显著性目标检测算法[J]. 光学学报, 2019, 39(9): 0915005. Xueli Xie, Chuanxiang Li, Xiaogang Yang, Jianxiang Xi. Salient Object Detection Algorithm Based on Dual-Attention Recurrent Convolution[J]. Acta Optica Sinica, 2019, 39(9): 0915005.