Adaptive Scaling for U-Net in Time Series Classification

Abstract

Convolutional Neural Networks such as U-Net are recently getting popular among researchers in many applications, such as Biomedical Image Segmentation. U-Net is one of the popular deep Convolutional Neural Networks which first contracts the input image using pooling layers and then upscales the feature maps before classifying them. In this paper, we explore the performance of adaptive scaling for U-Net in time series classification. Also, to improve performance, we extract features from the trained U-Net model and use ensemble deep Random Vector Functional Link (edRVFL) to classify them. Experiments on 55 large UCR datasets reveal that adaptive scaling improves the performance of U-Net in time series classification. Also, using edRVFL on extracted features from the trained U-Net model enhances performance. Consequently, our U-Net-edRVFL classifier outperforms other time series classification methods. 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.