A Multi-Modal RGB-D Capture Detection Network with Joint Spatial Domain and Frequency Domain Alignment

In modern vision-based robot grasping inspection systems, the effective fusion of color image and depth image is very important to improve the accuracy and robustness of the system. Traditional grasp detection methods often rely only on pixel alignment on the spatial domain, which can limit their effectiveness in complex environments. In this paper, a two-domain alignment mechanism is introduced, which utilizes both spatial domain and frequency domain information to achieve more robust and accurate grasping prediction (FSGNet). The proposed network architecture uses a two-branch structure to process the features of color image and depth image respectively, and uses the Adaptive edge-Aware Attention (AEAA) Module to enhance the model's attention to Edge features. After processing edge features, frequency domain information fusion is carried out by frequency-domain perceptive attention module (FA) and weighted fusion module (WF) based on depth-separable convolution. In this paper, comparative experiments are carried out on the Jacquard dataset and the self-built complex background object dataset to prove the effectiveness and robustness of FSGNet.