Abstract: Aiming at the inadequacy of the existing grasping detection methods that cannot effectively perceive sparse and weak features, thus leading to the degradation of robot grasping detection performance in low-light environments, robotic grasp detection with feature fusion of spatial and Fourier domains under low-light environments is proposed. The proposed model utilizes an encoder-decoder architecture as its backbone. During the fusion of deep and shallow features within the network, a hybrid feature extraction approach combining the spatial and Fourier domains is employed. In the spatial domain, global contextual information is captured using strip convolutions applied in horizontal and vertical directions, enabling the extraction of information critical to the grasp detection task. In the Fourier domain, image details and texture features are enhanced by independently modulating amplitude and phase components. Furthermore, a R-CoA attention module is incorporated to effectively balance global and local image information. This module encodes the relative positional relationships of image rows and columns, allowing the model to emphasize positional information pertinent to grasp tasks. The performance of the method is validated on three datasets: low-light Cornell/Jacquard datasets and the proposed low-light C-Cornell dataset. The proposed method ach ieves highest accuracy rates of 96.62%, 92.01%, and 95.50%, respectively. Specifically, on the low-light Cornell dataset (Gaussian noise and gamma=1.5), the proposed method improves by 2.24 percentage points and 1.12 percentage points, respectively, compared to GR-ConvNetv2, Se-ResUnet. The proposed method can effectively improve the robustness and accuracy of grasp detection in low-light environments, which provides support for grasp tasks under low-light conditions.

Key words: robot, grasp detection, spatial-fourier domain, attentional mechanism, deep neural network

摘要： 针对现有抓取检测方法无法有效感知稀疏，微弱特征，从而导致低光环境下机器人抓取检测性能下降的不足，提出一种融合空间-傅里叶域信息的机器人低光环境抓取检测方法。 首先，该方法中骨干网络采用编-解码器结构，在网络深层特征与浅层特征融合过程中，进行空间域-傅里叶域的特征提取。具体地，在空间域中通过水平和垂直方向的条带卷积捕获全局上下文信息，提取对抓取检测任务敏感的特征；在傅里叶域中通过分别调整振幅和相位，实现对图像细节和纹理特征的恢复。其次，引入R-CoA注意力模块平衡图像全局与局部信息，并对图像进行行、列相对位置编码以强化与抓取任务相关的位置信息。最后，在低光Cornell、低光Jacquard以及所构建的低光C-Cornell数据集上分别进行验证，所提低光抓取检测方法最高准确率分别达到96.62%、92.01%和95.50%。在低光Cornell数据集（高斯噪声且伽马值为1.5）上，与GR-ConvNetv2、Se-ResUnet相比，所提方法分别提升2.24个百分点和1.12个百分点。 所提方法能够在低光环境下有效提升抓取检测的鲁棒性和准确性，为机器人在低光照条件下的抓取任务提供支持。

关键词: 机器人, 抓取检测, 空间-傅里叶域, 注意力机制, 深度神经网络