Random Cropping Ensemble Neural Network for Image Classification in a Robotic Arm Grasping System
Bo Cheng, Wanyin Wu, Dapeng Tao, Shibo Mei, Ting Mao, Jun Cheng
Abstract
Robotic arms are expected to perform increasingly complex and intelligent tasks in industrial applications. These functions depend, to a large extent, on successful image classification. However, conventional image classification methods perform less well on randomly placed parts where there is uneven illumination in the industrial environment. Through detailed analysis of images taken in an industrial grasping environment, here, we propose a weighted ensemble neural network that can effectively overcome the impact of uneven placement and illumination over an entire image. For an image with overlapping parts, we combine ensemble learning with random cropping and develop the random cropping ensemble neural network (RCE-NN). RCE-NN has two main components: 1) random cropping helps to select different subimages in the image and 2) the weighted ensemble handles the results of each subimage classification to reduce subimage noise. In this article, the uncertainty of the class probability is considered, and therefore, the entropy of the subimage is used to predict the weight. We conduct experiments on our new FIST-Parts10 data set to verify the effectiveness of RCE-NN for image classification.