A two-stage multi-fidelity design optimization for K-mer-based pattern recognition (KPR) in image processing

Pattern recognition has been widely used in various applications of image processing. It is used to extract meaningful image features from the given image samples and to build classification systems with the intelligence of human recognition. Convolutional Neural Network (CNN) [1] has been one of the most popular and widely used methods for image pattern recognition applications. However, CNN was known not to be rotation-invariant to image patterns. It usually required a larger amount of training image dataset with greater variations in positions and orientations, or additional numerical treatments of spatial transformations [2]. On the other hand, K-mer-based Pattern Recognition (KPR) [3] has been developed to apply an unique way of rotation-invariant sampling to the inspected image pattern and analyze the frequency of the captured pattern features. A classification system was then built based on the Kmer frequency for the desired pattern recognition. In this paper, a series of tests and verifications of the KPR was done. It was found that finding the appropriate design parameters of the KPR for a specific application of image pattern recognition could be costly. A two-stage multi-fidelity design optimization was utilized to improve the efficiency of finding the parameters of KPR. In each iteration of the multi-fidelity design optimization procedure, the first stage was to evaluate the accuracy and efficiency of the design parameters in the K-mer-based pattern classification based on a full set of the given images. The second stage was to find a newer set of design parameters that performed the best based on a smaller set of images, which provided a classification set with lower fidelity than the original one. As a result, the proposed strategy of the multi-fidelity design optimization was more efficient than finding the optimal design parameters based on the full set of the given images.