A lightweight object grasping network using GhostNet

Object grasping is a very challenging problem in computer vision and robotics. Existing algorithms generally have a large number of training parameters, which lead to long training times and require high performance facilities. In this paper, we present a lightweight neural network to solve the problem of object grasping. Our network is able to generate grasps at real-time speeds (∼30ms), thus can be used on mobile devices. The main idea of GhostNet is to reduce the number of parameters by generating feature maps from each other in the process of convolution. We adopt this idea and apply it on the deconvolution process. Besides, we construct the lightweight grasp network based on these two processes. A lot of experiments on grasping datasets demonstrate that our network performs well. We achieve accuracy of 94% on Cornell grasp dataset and 91.8% on Jacquard dataset. At the same time, compared to traditional models, our model only requires 15% of the number of parameters and 47% of training time.