FPGA圖像處理-Sobel邊緣檢測原理

第第頁FPGA圖像處理-Sobel邊緣檢測原理因?yàn)樵谧?*3卷積的時(shí)候，圖像大小會(huì)變小，具體計(jì)算公式如下

其中O是輸出特征圖的大小，I是輸入特征圖的大小，P是Padding的大小，K是卷積核的大小，S是指Stride的大小，當(dāng)K的值是3，P的值是1，S的值也是1，的時(shí)候O的值和I的值相等。

為了保持輸出圖像的大小在經(jīng)過卷積后和輸入的大小一樣，我們需要進(jìn)行Padding操作，在這里我采用了復(fù)制周圍一圈的方式來完成。

采用（python）完成Sobel（算法）的參考模型

imp（or）tcv2ascvimportnumpyasnpimg=cv.imre（ad）(r"G:shiyan（IDc）.jpg")img_gray=cv.cvtColor(img,cv.COLOR_RGB2GRAY)h,w=img_gray.shapeimg_padding=np.ze（ros）((h+2,w+2),np.uint8)img_padding[1:h+1,1:w+1]=img_grayimg_padding[0:1,1:w+1]=img_gray[0:1,:]img_padding[h+1:h+2,1:w+1]=img_gray[h-1:h,:]img_padding[:,0:1]=img_padding[:,1:2]img_padding[:,w+1:w+2]=img_padding[:,w:w+1]th=200sobel_（rf）=np.zeros((h,w),np.uint8)foriinrange(1,h):forjinrange(1,w):gx1=img_padding[i-1][j+1]+2*img_padding[i][j+1]+img_padding[i+1][j+1]gx2=img_padding[i-1][j-1]+2*img_padding[i][j-1]+img_padding[i+1][j-1]gy1=img_padding[i-1][j-1]+2*img_padding[i-1][j]+img_padding[i-1][j+1]gy2=img_padding[i+1][j-1]+2*img_padding[i+1][j]+img_padding[i+1][j+1]gx=abs(gx1-gx2)gy=abs(gy1-gy2)ifgx+gy>th:sobel_rf[i-1][j-1]=255else:sobel_rf[i-1][j-1]=0cv.imshow("sobel_rf",sobel_rf)cv.imshow("src",img_gray)cv.w（ai）tKey()cv.destroyAllWindows()

根據(jù)算法模型完成HDL：提供SpinalHDL源碼

importspinal.core._importspinal.lib._classSobel(th:Int,imageColNum:Int,imageRowNum:Int)ex（te）ndsComponent{valio=newBundle{valdataIn=slave(ImageStream(8,imageColNum,imageRowNum,1))valdataOut=master(ImageStream(8,imageColNum,imageRowNum,1))}noIoPrefix()valgenMatrix=newGenMatrix(scala.math.pow(2,log2Up(imageColNum)).toInt,imageColNum,imageRowNum)genMatrix.io.dataInio.dataInvalgenMatrixOut=ImageStream(8,imageColNum,imageRowNum,9)genMatrixOut:=genMatrix.io.dataOutvalGX1=RegNext(genMatrixOut.data(0).asUInt+^(genMatrixOut.data(1)##B"1'b0").asUInt+^genMatrixOut.data(2).asUInt)valGX2=RegNext(genMatrixOut.data(6).asUInt+^(genMatrixOut.data(7)##B"1'b0").asUInt+^genMatrixOut.data(8).asUInt)valGX=Reg(UInt(11bits))valGY1=RegNext(genMatrixOut.data(6).asUInt+^(genMatrixOut.data(3)##B"1'b0").asUInt+^genMatrixOut.data(0).asUInt)valGY2=RegNext(genMatrixOut.data(8).asUInt+^(genMatrixOut.data(5)##B"1'b0").asUInt+^genMatrixOut.data(2).asUInt)valGY=Reg(UInt(11bits))when(GX1>GX2){GX:=GX1-GX2}otherwise{GX:=GX2-GX1}when(GY1>GY2){GY:=GY1-GY2}otherwise{GY:=GY2-GY1}valG=RegNext(GX+GY)valsobelOut=Reg(Bits(8bits))when(G>th){sobelOut:=255}otherwise{sobelOut:=0}io.dataOut.data(0):=sobelOutio.dataOut.row:=Delay(genMatrixOut.row,4)io.dataOut.col:=Delay(genMatrixOut.col,4)io.dataOut.c.hsync:=Delay(genMatrixOut.c.hsync,4,init=False)io.dataOut.c.vsync:=Delay(genMatrixOut.c.vsync,4,init=False)io.dataOut.c.de:=Delay(genMatrixOut.c.de,4,init=False)}objectSobelextendsApp{SpinalConfig().generate（Verilog）(newSobel(200,640,480))}

（仿真）代碼：

importspinal.lib._importspinal.core._importspinal.core.（sim）._importscala.collec（ti）on.mutable.Queueimport（java）.io.FileOutputStreamimportscala.io.SourceclasstbSobelC(th:Int)extendsSobel(th,430,430){varsrc=Array[String]()vardestDut=Array[String]()vardestRef=Array[String]()//varsrcLen=0varwidth=Array[Int]()varhigh=Array[Int]()valdutData=Queue[Int]()valrefData=Queue[Int]()varframeLen=0definit(srcFile:Array[String],destDu（tF）ile:Array[String],destRefFile:Array[String],imgShape:Array[(Int,Int)])={clockDomain.forkStimulus(10)io.dataIn.data(0)#=0io.dataIn.row#=0io.dataIn.col#=0src=srcFiledestDut=destDutFiledestRef=destRefFileio.dataIn.c.de#=falseio.dataIn.c.vsync#=falseio.dataIn.c.hsync#=falseframeLen=src.lengthwidth=imgShape.map(i=>i._1)high=imgShape.map(i=>i._2)clockDomain.waitSampling(10)}defframe(src:String,width:Int,high:Int)={valsrcFile=Source.fromFile(src)valsrcData=srcFile.getLines()varcolCnt=0varrowCnt=0io.dataIn.row#=widthio.dataIn.col#=highio.dataIn.c.de#=falseio.dataIn.c.vsync#=falseio.dataIn.c.hsync#=falseclockDomain.waitSampling(20)while(srcData.hasNext){valdata=srcData.next()io.dataIn.data(0)#=data.toIntio.dataIn.c.de#=trueif(colCnt==0&&rowCnt==0){io.dataIn.c.vsync#=trueprintln("xx")}else{io.dataIn.c.vsync#=false}if(colCnt==width-1&&rowCnt==high-1){clockDomain.waitSampling(1)io.dataIn.c.de#=falseclockDomain.waitSampling(200)}if(colCnt==0){io.dataIn.c.hsync#=true}else{io.dataIn.c.hsync#=false}if(colCnt==width-1&&rowCnt!=high-1){clockDomain.waitSampling(1)io.dataIn.c.de#=falseclockDomain.waitSampling(20)}if(colCnt==width-1){colCnt=0if(rowCnt==high-1){