Human inspired pattern recognition via local invariant features.

摘要

Vision is increasingly becoming a vital element in the manufacturing industry. As complex as it already is, vision is becoming even more difficult to implement in a pattern recognition environment as it converges toward the level of what humans visualize. Relevant brain work technologies are allowing vision systems to add capability and tasks that were long reserved for humans. The ability to recognize patterns like humans do is a good goal in terms of performance metrics for manufacturing activities. To achieve this goal, we created a neural network that achieves pattern recognition analogous to the human visual cortex using high quality keypoints by optimizing the scale space and pairing keypoints with edges as input into the model.;This research uses the Taguchi Design of Experiments approach to find optimal values for the SIFT parameters with respect to finding correct matches between images that vary in rotation and scale. The approach used the Taguchi L18 matrix to determine the optimal parameter set. The performance obtained from SIFT using the optimal solution was compared with the performance from the original SIFT algorithm parameters. It is shown that correct matches between an original image and a scaled, rotated, or scaled and rotated version of that image improves by 17% using the optimal values of the SIFT.;A human inspired approach was used to create a CMAC based neural network capable of pattern recognition. A comparison of a 3 object, 30 object, and 50 object scenes were examined using edge and optimized SIFT based features as inputs and produced extensible results from 3 to 50 objects based on classification performance. The classification results prove that we achieve a high level of pattern recognition that ranged from 96.1% to 100% for objects under consideration. The result is a pattern recognition model capable of locally based classification based on invariant information without the need for sets of information that include input sensory data that is not necessarily invariant (background data, raw pixel data, viewpoint angles) that global models rely on in pattern recognition.