以輕量化卷積神經網路為核心之自動抄錶系統

Abstract

隨著物聯網技術的蓬勃發展，政府逐漸淘汰了傳統電錶，開始了智能電錶的時代。然而，更換智能電錶的價格昂貴且面臨通訊不良等問題，導致智能電錶佈建緩慢，我們的想法是開發一種低成本的解決方案，該解決方案使用帶有攝影鏡頭的邊緣設備自動辨識傳統電錶，然後將辨識的值上傳到雲端。過去已有研究通過傳統的圖像分割方法自動讀取錶盤，但是由於傳統的電錶大多設置在遮蔽性高、光線昏暗、灰塵多的環境中，因此對於不清晰的電錶圖像，傳統方法難以獲得良好的辨識結果。在本文中，我們提出了一種基於輕量化卷積神經網路的自動讀錶器並實現在邊緣設備上，為了減輕佈建難度和提高錶盤辨識的準確率，我們所提出的錶盤讀取器具有自動調整傾斜錶盤圖像的能力。實驗結果顯示，相較於其他相關方法，所提出的輕量化卷積神經網路在分割錯誤，誤報和運行時間方面取得了顯著改善。

With the vigorous development of the Internet of Things technology, the government has gradually phased out the traditional meter and began the era of smart meters. However, the replacement of smart meters is expensive and the yield is too low, which has led to the slow deployment of smart meters. Our idea is to develop a low-cost alternative solution that uses an edge device with a camera to automatically identify traditional electric dial meters, and then uploads the identified value to cloud servers. In the past, there have been studies to automatically read dial meters through traditional image segmentation methods. However, because traditional electric meters are mostly set in an environment with high concealment, dim light, and dirt, it is difficult for traditional methods to obtain good identification results for unclear meter images. In this thesis, we propose a cost-effective automatic dial meter reader with a lightweight convolutional neural network on edge devices. In order to easily deploy and improve the accuracy of dial meter recognition, the proposed meter reader has the ability to automatically adjust tilt meter images. Experimental results show that the proposed lightweight convolutional neural network achieves significant improvements in segmentation errors, false positives, and elapsed time compared with the relative approaches.