基於田口法與FPGA平行運算辨識物件顏色之影像處理

Object Recognition of Image Processing Based on Taguchi Method and FPGA Parallel Computing

李彥儀(Yan-Yi Li)

指導教授 ： 張文陽

國立虎尾科技大學/工程學院/機械與電腦輔助工程系碩士班/碩士(2015年)

https://doi.org/10.6827/NFU.2015.00030

摘要

本論文研究主要目的為運用田口統計分析與FPGA(Field programmable gate array)平行運算平台，進行分析不同物件顏色之影像辨識與處理。FPGA硬體平台使用DE2-115發展板，其內含處理器為Altera Cyclone IV的FPGA晶片，並利用其平行機制的特性，於每時脈循環可同時完成多重平行運算作業，因此超越數位訊號處理器(Digital signal processor，簡稱DSP)的計算速度，硬體層級控制Input/Output(I/O)與其他硬體比較更可縮短響應時間。物件顏色閥值方面則是以Red/Green/Blue(RGB)空間為色域基礎，並將整體影像轉為二值化以利於分離目標色彩與其它色彩，本實驗以Red、Green及Blue的物件作為辨識目標，並透過各物件有效像素色彩值的統計，接著使用田口品質方法將辨識閥值上調與下調一定的裕度後，排列成全因子的形式並再度重新進行最佳化色彩辨識，以探討是否提高物件辨識的精準度。
最後嘗試將全因子的排列方式寫入FPGA平台，並將閥值的有效像素值顯示於七段顯示器上，與依照色碼表所找出的手動閥值辨識結果做比對及驗證，且使用田口品質方法進行S/N比以及變異數分析的運算，找出各因子所造成的貢獻度。所有影像部分皆以VGA或是VEEK-MT作為輸出平台；程式方面使用Quartus II 11.1進行編譯，而硬體描述語言部份則是使用Verilog HDL的方式來實現以上的方法；田口品質方法以望大型來計算有效像素值的S/N信號雜音比。變異數分析紅色物件的辨識結果，最佳化組合為R1G3B2，主要趨勢影響最大的是B因子，貢獻度也是B因子較多為78.73824%，誤差比例為8.472%；綠色物件的最佳化組合為R2G1B3，主要趨勢影響最大的是G因子，其貢獻度也最多為62.44788%，誤差比例為8.626%；藍色物件的最佳化組合為R3G3B3，影響主要趨勢最高的也是G因子，貢獻度為36.40271%，但其誤差高達35.018%

關鍵字

FPGA ； 平行運算 ； 影像處理 ； 田口品質方法 ； 最佳化分析

並列摘要

This is the study of object recognition using image processing based on Taguchi method and FPGA parallel computing. The hardware platform of image processing uses the DE2-115 development board with chip-Altera Cyclone IV. The program compiler of FPGA is the Quartus II 11.1 with the hardware description language in Verilog HDL. As we know that the FPGA model can simultaneously execute the image processing and the calculation speed of the image processing is quicker than DSP. The response speed of the FPGA hardware is faster than the results from other hardware. The color threshold setting for image recognition is based on RGB space with a binary image. The target color and background other color are separated into red, green and blue objects. The Taguchi method used the thresholds for these pixel values. These data are arranged for full factorial design (FFD) reversed into the form of optimized color identification, in order to explore the possibility of improving on identification accuracy.
The color image recognitions of the objects are estimated using ideal and optimal threshold methods. The ideal threshold is a constant value. FFD is used to perform the optimal threshold of color image recognition in FPGA platform. The color code table is used to manually identify optimal threshold levels. The color areas of the object are then recognized based on the optimal and ideal threshold values. Each threshold value produces effective pixels to be shown on the seven-segment display. Finally, the color areas of the FFD is compared and verified from ideal threshold method. FFD threshold identifications are used to estimate S/N ratios and ANOVA (analysis of variance) calculations which can evaluate the contribution of each factor. All images of object recognition are presented as VGA or VEEK-MT. The effective pixel values of the S/N signal to noise ratios were calculated by Taguchi method of “larger the better”. The ANOVA is analyzed for various object identifications.
The result showed that the optimal level of the red object for control factors of the red, green, and blue are level 1, level 3, and level 2, respectively. The level 1, level 3, and level 2 intensities of the red, green, and blue are 208~837, 16.5~217.8, and 15~246 , respectively. The major control factor of the red object is blue factor with a contribution factor of 78.74% and the error ratio is 8.47%. The optimal level of the green object for control factors of the red, green, and blue are level 2, level 1, and level 3, respectively. The level 2, level 1, and level 3 intensities of the red, green, and blue are 30~208, 151.2~905.4, and 33~228.8, respectively. The major control factor of the red object is blue factor with a contribution factor of 62.45% and the error ratio is 8.63%. The optimal level of the blue object for control factors of the red, green, and blue are level 3, level 3, and level 3, respectively. The level 3, level 3, and level 3 intensities of the red, green, and blue are 34.1~195.8, 34.1~204.6, and 114.4~1023, respectively. The major control factor of the red object is blue factor with a contribution factor of 36.4% and the error ratio is 35.02%. Finally, the recognition rates of red, blue, and green objects based on optimal threshold methods are 93.68%, 88.21, and 92.25%, respectively. However, the recognition rates of red, blue, and green based on optimal threshold methods are 98.62%, 96.21, and 94.86%, respectively.

並列關鍵字

FPGA ； Parallel computing ； Image processing ； Taguchi methods ； Optimization

參考文獻

[7]蔡治平，2013， “影像視覺之光譜照度分析與物件特徵辨識” ，國立虎尾科技大學，機械與電腦輔助工程研究所碩士論文。

[18]黎公海，2014， “以FPGA實現即時手追蹤系統” ，國立虎尾科技大學，資訊工程研究所碩士論文。

[9]繆孝君，2010， “以FPGA實現動態物體追蹤系統” ，國立虎尾科技大學，資訊工程研究所碩士論文。

[19]楊承翰，2014， “基於FPGA之物件影像自動化辨識與檢測系統” ，國立虎尾科技大學，機械與電腦輔助工程研究所碩士論文。

[17]蔣國祥，2013， “以FPGA為基礎實現且應用於機械手臂之影像辨識系統” ，國立成功大學，電機工程研究所碩士論文。