基於深度學習的物件偵測演算法之探討

Investigation of Deep Learning Based Object Detection Algorithms

楊鎔謙(RONG-QIAN YANG)

指導教授 ： 洪文斌

淡江大學/工學院/資訊工程學系資訊網路與多媒體碩士班/碩士(2020年)

https://doi.org/10.6846/TKU.2020.00877

摘要

本論文討論如何解決礙子異常放電的問題，主要以如何準確定位礙子、及判斷礙子是否放電為主要討論方向。首先透過放電偵測儀搭配空拍機及固定的攝影機擷取影片，本論文有兩個實驗，實驗一為使用捲積神經網路(Convolutional Neural Network, CNN)來定位礙子，實驗二利用YOLOv3來定位礙子。利用CNN訓練出來的準確率約為0.95至0.98之間，由於利用CNN來定位礙子，最後會在判斷放電區域是否在礙子上出現問題，為追求更精確的實驗結果，後來透過YOLOv3定位方法直接定位礙子設備的正確位置，判斷紅色像素的數量是否超過標準值，此方法的定位效果佳，也能實作出系統。本論文利用傳統檢測的標準(mAP)分析YOLOv3的定位結果，模組的信心分數平均約莫在0.98～1.00之間，平均密度精度為(mAP)98.68%。

關鍵字

深度學習 ； 卷積神經網路 ； YOLO ； mAP ； IOU

並列摘要

This article discusses how to solve the problem of abnormal discharge of Insulators, mainly focusing on how to accurately locate the Insulator and determine whether the Insulator is discharged. First, an infrared detector with an aerial camera and a fixed camera is used to capture the video. There are two experiments in this article. The first experiment is to use Convolutional Neural Network (CNN) to locate Insulators, and the second experiment is to use YOLOv3. Look for Insulators . The accuracy of CNN training is about 0.95 to 0.98. Since CNN is used to locate Insulators , it will finally determine whether there is a problem with the emission area. In order to obtain more accurate experimental results, the YOLOv3 positioning method was later used to directly locate the correct position of the blocking sub-device and determine whether the number of red pixels exceeds the standard value. This method has good positioning effect and can also be implemented as a system. This article uses traditional detection standards (mAP) to analyze the positioning results of YOLOv3. The confidence score of this module is about 0.98-1.00 on average, and the average density accuracy (mAP) is 98.68%.

並列關鍵字

Deep Learning ； CNN ； YOLO ； mAP ； IOU

參考文獻

[1] B. Knyazev, R. Shvetsov, N. Efremova and A.Kuharenko, “Leveraging Large Face Recognition Data for Emotion Classification,” In IEEE International Conference on Automatic Face & Gesture Recognition, 2018.

[2] S. Sadhu, R. Li, and H. Hermansky, “M-vectors: Sub-band Based Energy Modulation Features for Multi-stream Automatic Speech Recognition,” In IEEE International Conference on Acoustics, Speech and Signal Processing, 2019.

[3] D. C. Ciresan, U. Meier, J. Masci, L. M. Gambardella, and J. Schmidhuber, “High Performance Convolution Neural Networks For Images Classification,” In Twenty-second International Joint Conference on Artificial Intelligence, 2011.

[4] R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich Feature Hierarchies For Accurate Object Detection And Semantic Segmentation,” In CVPR, 2014.

[5] P. Felzenswalb and D. Huttenlocher, “Efficient Graph Based Image Segmentation,” In International Journal on Computer Vision, pp.169-181, 2004.