Normalization of data for training and analysis by the MaskRCNN model using the k-means method for a smart refrigerator’s computer vision

Abstract

The authors propose to use the k-means method for semantic image segmentation in the artificial vision of a smart refrigerator. The authors have developed a new two-tier architecture for the semantic segmentation system. In the proposed architecture, various sets of image contrast optimization settings are used to classify pixels belonging to fragments of the studied image. Experiments have shown that the proposed method has decisive advantages over existing works. The k-means method was used to cluster pixels directly into semantic groups. The clustered data is used to solve the semantic segmentation problem. For the correct selection of the number of centroids of the k-means method, the mean shift method was used. Extensive experiments with the traditional k-means algorithm and mean shift method have shown the advantage of the proposed approach in accuracy and performance.