Generating training data for neural‐network‐based pose recognition in parts feeding Available to Purchase

The capability of an artificial neural network to determine part pose by processing image data from the silhouette of a back‐lit part has been established in recently reported work. The chief benefit of this new approach is simplicity of training, which is important for flexible automated parts feeders. The objective of the work presented herein is to develop an effective and efficient method for determining the position and orientation of the parts to be used in training the neural network. Candidate methods were used to create sets of training data containing different numbers of images taken of each part in different patterns of position and orientation. For each set of training data, the neural network was trained and its pose recognition performance was empirically evaluated. Based on these empirical results, a method for generating training data is reported that ensures accurate performance of the trained neural network while requiring only a minimum amount of training data.