Neural-network-based pose recognition in flexible parts feeders: preliminary results for reduced training data

Artificial neural networks have been shown to be an effective approach for recognizing part pose in vision-based, flexible parts feeders. In this approach the neural network processes simple image data from the silhouette of the back-lit part to determine part pose. In an extensive empirical evaluation in which 500 images per part were used to calibrate a neural network for each of six parts from a desktop printer, these neural networks performed flawlessly. The chief benefit of this approach is simplicity of training, which is important for flexible automated parts feeders. The long-term 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. An effective method must ensure satisfactory performance of the trained neural network, whereas an efficient method requires that as few images as possible be used in training the neural network. The first step toward reducing the amount of training data required was to determine whether the high level of pose recognition performance achieved in the baseline study can be achieved with networks trained with less calibration data. Accordingly, sets of training data were generated using images taken of each part in a specified regular pattern of positions and orientations. Across the six parts studied, an average of 264 images per part were used in these data sets. Based on an empirical evaluation of pose determination performance of the neural network calibrated with these sets of training data, flawless pose recognition performance was achieved.