In aircraft manufacturing, effective management and storage of aircraft parts are essential for enhancing production efficiency. Considering the challenge of fully deploying high-end hardware in industrial settings, this paper introduces a attention-based lightweight network for aircraft part grasping detection method for achieving high precision and rapid robotic grasping. The network improves detection accuracy through feature fusion and attention mechanisms. Specifically, within the attention mechanism module, depthwise separable convolution is used in place of fully connected layers to reduce the number of parameters. The network employs depthwise separable convolution and max pooling to enhance features and uses instance normalization to accelerate the network learning speed. Furthermore, a novel Log-Cosh loss function is introduced, which stabilizes gradients with an adaptive constant. Quantitative experimental results are compared with other methods, showing 98.9% accuracy, a speed of 25.0ms, and a parameter volume of 0.407M for both RGB and RGB-D images. In qualitative tests, the confidence for single grasping exceeds 0.78, and the average confidence for multiple graspings is 0.77. In PyBullet simulation, the grasping success rate for 40 different objects is 80.10%.
 

Grasping detection, Lightweight network, Instance segmentation, AI applications in engineering