Grasp Detection Algorithm Based on CSP-ResNet

Aiming at the problem of low detection accuracy of grasping detection algorithm based on RGB information as input, this paper proposes a grasping detection algorithm based on CSP-ResNet to improve the detection accuracy of the algorithm. Aiming at the problem that the chessboard effect will be produced when the transposed convolution restores the image fractional variability, which will affect the prediction accuracy of the model, this paper designs the fusion of the nearest neighbor interpolation upsampling and the transposed convolution upsampling to restore the resolution of the feature map. To alleviate the checkerboard effect produced by the transposed convolution. Namely FCG-Net (Fuse CSPS Grasp Net). This method takes RGB information as input, first extracts feature through cubic convolutional neural network, uses CSPS-ResNet structure to improve the learning ability of the network, extracts effective feature layers, and finally adopts the method of nearest neighbor interpolation and transposed convolution upsampling fusion. The resolution of the recovered image is decoded to obtain the optimal grasping pose. In order to verify the effectiveness of the algorithm in this paper, based on the Cornell grasping dataset, grasping detection and physical grasping experiments were carried out. In the data set format divided by Object-wise, the accuracy of grasp detection reaches 97.8%. The image-wise partitioned dataset achieves 95.5% accuracy in grasping poses.