深度學習影像偵測結果分析工具

Abstract

中文摘要 對於視覺神經網路的蓬勃發展，越來越龐大的運算需要移植到邊緣運算平台，需要更低的能耗表現，常使用量化(Quantize)的技術來降低資料寬度，這點讓資料搬運、儲存與運算的成本相較於單精度浮點數的成本更低。但是從中衍生的就是精確度損失的問題。 對於以物件框輸出作為偵測結果的神經網路，通常以該神經網路對於相同的Ground Truth資料集進行測試所得出的mAP (mean Average Precision)來評估神經網路的準確性。mAP的確能表示該神經網路的準確度，但是以實際應用上，仍然需要直接檢視結果來判斷。 這次研究是希望單純以物件框的在原始圖片中的位置來觀察神經網路的偵測結果，也就是更方便的比較不同神經網路和實現流程在實務上的成果。並且在自己的實驗中，需要比較相同測試資料在不同資料寬度和網路架構的偵測結果。或能透過這次做出的工具，在之後的深度學習網路的訓練中，增強神經網路訓練的不足或對結構做調整。 首先我先訓練出幾組權重，並透過Quantization將其轉換為int8的資料格式，在不同平台進行物件偵測，並取得各自的偵測結果。接著透過PyQt5設計一個UI來讓我想要觀察的資料能夠更容易的直接呈現。最後，這裡做出了一個能夠比較相同Ground Truth透過不同神經網路實現流程的偵測結果的工具。

Abstract Because of the flourishing development of visual neural networks, in order to transplant more and more large calculations to the edge computing platform, lower energy consumption performance is required. And quantization is often used to reduce the data width. This makes the cost of data handling, storage, and calculations lower than the cost of single-precision floating-point data. But what derives from it is loss of detection accuracy. For convolution neural networks(CNNs) that use bounding boxes as the detection result, the accuracy of the CNNs are usually evaluated by mAP (mean Average Precision) obtained by testing the CNNs on the same ground truth data set. This can indeed indicate the accuracy of the CNNs, but in practical applications, it is still necessary to directly inspect the results to judge. This research hopes to observe the detection results of the CNNs by checking the position of the bounding boxes in the image, that is, it is more convenient to compare the practical results of different CNNs and implementation processes. And in my own experiment, I need to compare the detection results of the same test data in different data widths and network architectures. Perhaps through the tools made this time, in the subsequent deep learning network training, the deficiencies of neural network training can be enhanced or the structure can be adjusted. First, I train a few sets of CNN weights, and convert them to int8 data format through quantification, perform object detection on different platforms, and obtain their respective detection results. Then design a UI through PyQt5 to make the data I want to observe more easily and directly presented. Finally, here is a tool that can compare the detection results of the same ground truth through different neural network implementation processes.