以卷積類神經網路應用於大尺度粒子影像測速之研究

Abstract

河川流量資料是水利工程設計與水資源規劃中最關鍵的因子，近 20 年來使用非接觸式的方法量測水文資料已經獲得廣泛地應用，如大尺度粒子影像測速法(Large-Scale Particle Image Velocimetry, LSPIV)量測河道表面流速來估算流量的方法。然傳統粒子影像測速法所利用的圖像匹配法，僅考量圖像之強度分布的互相關性算法(Direct Cross-Correlation, DCC)作為匹配基準。就室內實驗而言，因為光照條件容易控制，使得該法可以獲得良好的速度場分布結果。但是在野外現場量測，因光照、陰影、粒子散佈、天然環境變數和人為條件均十分複雜且難以控制，使得拍攝的影像特徵不足，或是產生不可逆的雜訊，導致速度場的分析錯誤。因此以強度分布特徵作為影像辨識之判斷標準並無法適用於任何場域，特別是非接觸測量的方法要在現場應用來分析流量，勢必要開發一種可以考量亮度、幾何或是更抽象的特徵來進行辨識影像的方法提高速度場的量測精度。 卷積類神經網路(Convolutional Neural Network, CNN)其影像特徵擷取的準確度已遠超越傳統影像匹配法，是現今人工智慧影像辨識中應用最廣泛的技術。本研究利用 CNN 的特性，打造出適用於 LSPIV 應用場合的深度學習網路，來對散佈在水面上的粒子提取更多元的特徵進行影像辨識。本研究先以 Hamel-Oseen 渦流方程式產生兩種速度量級下的高密度、低密度，與背景擬真化之低密度三類影像基準集，分別模擬傳統室內實驗、疏鬆的顆粒分布與背景複雜的現場情況，並以向量相關係數來評估DCC和CNN在這些結果中的好壞。研究發現 CNN 對於背景亮度雜訊所造成的干擾抑制於 1%以內，在模擬現場情況的影像集中獲得了比DCC法還要穩健的表現。最後以 30 公尺長與 1 公尺寬的人造渠道進行實驗，使用聲學都普勒流速儀所量測的流速資料來評估 CNN 應用於LSPIV中之表現。結果顯示CNN與DCC法在流速量測表現極其接近，本研究證明CNN未來可以應用於現場的流速量測，同時為建立一套非接觸式流量量測儀建立基礎。

The discharge of river is the most important information in the water resource planning. Recently, one of the non-contact measurement techniques, “Large-Scale Particle Image Velocimetry (LSPIV)”, is widely used in measuring the surface flow velocity and estimating the discharge in the field. Conventionally, the direct cross-correlation algorithm (DCC) used in LSPIV considers the correlation of pixel intensity in the interrogation area as the characteristic. However, the condition is usually complex and not controllable in the field. These unfavorable factors not only probably impact the intensity patterns of the images you get, but also directly raise the error in matching interrogation area when applied on field. Only taking the intensity as the matching feature is not really reliable, we really need to consider more geometric characteristics in our matching technique. For the sake of improvement, in our research, we use “Convolutional Neural Network (CNN)”, which is very popular and powerful on the machine recognition and object detection, to take more characteristics on river surface into account, and build an appropriate deep learning network structure to extract and classify the features from the river surface. To validate the applicability of our method, we design three scenarios, which are high-density, low-density and low-density with non-uniform illuminated background, to simulate the most common circumstances of applying general PIV, LSPIV and field scene. Hence, we generate several particle image sets under those situations by using PIVlab. The particles on images are all based on the function of Lamb-Oseen Vortex rings, thus we have the true values of velocity field as our testing benchmarks, and use the vector correlation coefficient to check our results. We found that CNN which mitigating the impact of illumination below 1% can have more stable, and reliable results on the image sets that simulating the filed than the DCC which are seriously affected. Moreover, by having a experiment on the indoor channel, we proof that LSPIV with CNN can truly be used in reality, and have the same performance on the velocity field and flow rate with DCC method.