Data normalization for training and analysis of the MaskRCNN model using the k-means method for computer vision of smart refrigerator

Abstract

This paper proposes using the k-means method for the controlled adjustment of the training sample for semantic image segmentation in the artificial vision of a smart refrigerator. To solve this problem, a new two-stage architecture for computer vision is proposed. In the proposed architecture, various sets of settings for optimizing the contrast of images are used to classify pixels according to their belonging to fragments of the studied image. Extensive experimental evaluation shows that the proposed method has critical advantages over existing work. Firstly, the obtained pixel classes can be directly clustered into semantic groups using k-means. Secondly, the method can be used for additional training of artificial intelligence in solving the semantic segmentation problem. The developers propose an approach to the correct choice of the number k of centroids to obtain good quality clusters, which is difficult to determine at a high k value. To overcome the problem of initializing the k-means method, an incremental k-means clustering method is proposed, which improves the quality of clusters to reduce the sum of squared errors. Comprehensive experiments have been carried out compared to the traditional k-means algorithm and its new versions to evaluate the performance of the proposed method on synthetically generated datasets and some real-world datasets.