Pixel-Level Grasp Detection based on EfficientNet and Multi-scale Feature Fusion Network

Real-time performance and accuracy are pivotal for evaluating robotic grasping detection models. To enhance the detection precision while preserving real-time capabilities, this paper presents an EfficientGrasp-gdut network to identify the pixel-level robot grasping positions directly from RGB images. It incorporates one multi-scale convolutional attention (MSCA) module and one multi-scale feature fusion (MSFF) network to improve the robustness in unstructured environments based on EfficientNetV2. The experimental results show that it can attain up to 99.43% accuracy in grasp detection on both images and objects based on the Cornell dataset, with rate up to 26 frames per second (FPS). It has a significant advancement over some leading networks, such as AFFGA-Net and SE-ResUNet.