| 摘要: | 當前時代人工智慧發展迅速，許多傳統行業都逐漸從人工密集型尋求向人工智慧的轉型。在科技研究領域，深度學習的演算法不斷改進並在多個領域都得到應用。在目前關於儀表自動識別的研究中，使用傳統影像處理方法都會有很大的局限性，往往只能針對單一儀表，而且無法適應不同環境。本文結合深度學習與影像處理，嘗試用新方法辨識多種儀表並讀出儀表指數。
本文提出一套基於深度學習與影像處理之儀表偵測與讀數識別系統，其目的是對工廠中常見的儀表如圓形指針量表、方型指針量表，以及數字量表三種，進行自動偵測與讀數識別的工作。本文的主要工作流程如下，首先使用即時物件偵測網路YOLOv3從手機採集的圖像中定位儀表區域，使用影像前處理去除圖像的干擾資訊，便於後續圖像的處理；然後根據儀表的種類，使用HoughCircles和findContours函數，霍夫直線法和極座標法分別實現儀表表盤圓心定位，指針定位；再次使用即時物件偵測網路YOLOv3識別表盤的數字,根據其坐標找出離指針最近的兩組數字，結合指標的座標計算出數值，根據其坐標找出離指針最近的兩組數字，結合指標的座標計算出數值。
本論文使用即時物件偵測網路來迅速定位儀表位置，與傳統方法相比，能夠在不同的圖像條件下迅速準確地定位出各種類型和尺寸的表盤，具有很好的抗干擾性。整個系統適用於不同量程和不同種類的儀表識別，不需要給予任何儀表資訊，在提高了適用範圍的同時真正意義上實現了儀表的智能化讀數識別。
;With the rapid development of artificial intelligence, nowadays many traditional industries gradually from labor-intensive industry into artificial intelligence. In the field of scientific technology, deep learning algorithm has been improved and applied in many fields, especially in image processing and pattern recognition. Currently, traditional image processing methods have many limitations in the automatic instrument recognition and most of the proposed methods were designed for recognizing only one type of specific instrument meter. Therefore, this study proposes a novel method to recognize different instrument meters and read the scales pointed in the meters.
In this study, the main purpose is that a kind of intelligent system based on deep learning and image processing is designed to recognize common instruments such as circular meters, square meters and digital scale meters in the factory. First of all, we use object detection network YOLOv3 to locate instruments of images which are captured from mobile phone and camera. Then we use image processing technology to remove interference information of the images. Secondly, according to the type of instrument, HoughCircles and findContours functions are used to confirm the central position of the instrument, and hough line transform and polar coordinate method are used to realize the pointer position. In addition, YOLOv3 is applied again to recognize the number in order to find the closest number and the second closest number to the pointer, finally, we can calculate the value of the instrument.
This thesis uses object detection network to locate the instrument position. In comparison with the traditional methods, the proposed method can rapidly locate various types and sizes of the instrument in different images and has strong anti-interference. The system can be adapted to the different range and uneven scale of the instrument, which greatly improves the application range, and really realize the intelligent instrument recognition. |
