本研究針對不同土石流情境,分別以「機械視覺」理論研擬出影像前處理流程,建構土石流影像判識方法,並分析土石流特性,希冀可提供土石流監測之參考。 本研究使用Borland C++ 語言建構即時影像前處理與土石流即時判識程式,以現場及實驗室影像資料驗證判識程式,並研訂各相關門檻值。首先進行影像前處理,藉由「非均勻照明校正」、「亮度等值化」及「影像雜訊濾除」等三種影像強化方法,將擷取之影像清晰化並突顯主體。繼而進行土石流影像判識,針對土石流發生過程所產生之影像特徵進行研析,研擬出「特定物位移」、「土石流波前」、「影像紋理」與「多重濾波影像分類」等四種判識方法,以土石流發生紀錄片及室內土石流水槽模擬實驗影片進行影像判識模擬,驗證各判識方法相關參數值之合理範圍;最後分析土石特性參數,以影像分割與特徵擷取等方法估算土石流表面粒徑分布圖,並以礫石運動及最小臨近法推估礫石移動軌跡。 經由影像分析結果如下:在雨、霧與夜間等拍攝之不良影像,經影像前處理程式處理過後其主題凸顯清楚,有利於後續判識系統之進行;特定物位移可藉由「殘留機率」P,亦即像素變化百分比加以判識,程式判識結果與實際值相近,其平均誤差百分比為0.8%;波前判識可利用平均亮度之「擾動強度」以區別土石流或洪水;影像紋理參數經適當篩選後,在費雪區別分析與類神經網路分類模擬下,土石流判識正確率可達86%;而在土石流與洪水影像在多重濾波影像分類下其判識率可達八成;本文最後進行土石流特性分析,影像分析之礫石粒徑分佈曲線,與實際篩分析之粒徑分佈曲線比較明顯偏高,可以利用長短軸所建立關係式于以修正。
Aimed to the various scenarios, the ‘machine vision theory’ is applied to enhance the image quality and to build up the debris-flow image recognition and also to construct the size distribution curve of debris-flow. The characteristics of debris flow enhanced by the image processing techniques are to provide to the fundamental research of the debris flow detection. This study applied Borland C++ builder program language (BCB) to acquire image from CCD to PC at first. Thereafter, several image enhancement programs were developed to remove the noise in the images, which were taken from the filed in rain or fog. Three image-enhanced methods, the ‘non-uniform illumination’,‘intensity equalization’ and ‘noise-filtering’ were used to deal with the degraded image and to highlight the objects of interest. The ‘debris flow recognition program’ developed herein composed four functions,”detection of the specific objects moving”, “detection of the wavefront of debris flow”, “recognition of image texture” and“image recognition using filter banks”. The debris-flow recording videos and experimental videos at the flume test in laboratory are used to demonstrate the validity of the recognition algorithms. Then, the characteristics of debris flow were analyzed by using the image segregation and feature extraction techniques. The grain sizes of debris were then estimated and the size-frequency distribution curve can be drawn accordingly.The moving tracks of grain was estimated according to their moving directions and the nearest neighbor approach. Based on experimental data, the results can be summarized as follows. The image enhancement processing having done, the images became clear significantly and made the focal points stand out. The image enhancement may facilitate the follow-up debris-flow image recognition. The movement of the specific object can be determined by the percentage of image pixels of the specific object remaining in the region of interest. The experimental errors, comparing to the actual situations, are about 0.8%;The debris flow wave-front detection can discriminate debris-flow and flood with the fluctuation of their average intensity. With image textural and filter bank features,Fish’s linear discrimination model and bark-propagation neural network were be classified.The results show more than 86% and 80% in the processing case. Finally, the estimated grain sizes of gravels by using image analysis were greater than the real sizes of those. The correction parameters have found using the relations of short axes and long axes of grain.