图片质量评价参数FSIM计算代码-图片相似性参数FSIM源码-特征相似性FSIM-图片相似性评价参数FSIM计算

import numpy as np import tensorflow as tf import cv2 import math def lowpassfilter(sze,cutoff,n): if cutoff<0 or cutoff>0.5: print('cutoff frequency must be between 0 and 0.5') return if type(n)!=int or n<1: print('n must be an integer >= 1') return if type(sze)==int: rows=sze cols=sze else: rows=sze[0] cols=sze[1] if cols%2>0: xrange=np.arange(-(cols-1)/2,(cols-1)/2+1)/(cols-1) else: xrange=np.arange(-cols/2,cols/2)/cols if rows%2>0: yrange=np.arange(-(rows-1)/2,(rows-1)/2+1)/(rows-1) else: yrange=np.arange(-rows/2,rows/2)/rows x,y=np.meshgrid(xrange,yrange) radius=(x**2+y**2)**0.5 f=tf.signal.ifftshift(1.0/(1.0+(radius/cutoff)**(2*n))) return f.numpy() def phasecong2(im): nscale= 4 norient= 4 minWaveLength= 6 mult= 2 sigmaOnf= 0.55 dThetaOnSigma= 1.2 k= 2.0 epsilon= .0001 thetaSigma = math.pi/norient/dThetaOnSigma rows,cols=np.shape(im) imagefft=tf.signal.fft2d(im).numpy() EO=[[None for i in range(nscale)] for j in range(norient)] estMeanE2n=np.zeros((norient,)) ifftFilterArray=[None for x in range(nscale)] if cols%2>0: xrange=np.arange(-(cols-1)/2,(cols-1)/2+1)/(cols-1) else: xrange=np.arange(-cols/2,cols/2)/cols if rows%2>0: yrange=np.arange(-(rows-1)/2,(rows-1)/2+1)/(rows-1) else: yrange=np.arange(-rows/2,rows/2)/rows x,y=np.meshgrid(xrange,yrange) radius=(x**2+y**2)**0.5 theta=tf.math.atan2(-y,x).numpy() radius=tf.signal.ifftshift(radius).numpy() theta=tf.signal.ifftshift(theta).numpy() radius[0,0]=1 sintheta = tf.math.sin(theta).numpy() costheta = tf.math.cos(theta).numpy() lp=lowpassfilter([rows,cols],0.45,15) logGabor=[None for x in range(nscale)] for s in range(1,nscale+1): wavelength = minWaveLength*mult**(s-1) fo = 1.0/wavelength logGabor[s-1]=tf.math.exp((-(tf.math.log(radius/fo))**2) / tf.cast(2 * tf.math.log(sigmaOnf)**2,dtype=tf.float64)).numpy() logGabor[s-1]=logGabor[s-1]*lp logGabor[s-1][0,0]=0 spread=[None for x in range(nscale)] for o in range(1,norient+1): angl = (o-1)*math.pi/norient ds = sintheta * math.cos(angl) - costheta * math.sin(angl) dc = costheta * math.cos(angl) + sintheta * math.sin(angl) dtheta = tf.math.abs(tf.math.atan2(ds,dc)).numpy() spread[o-1] = tf.math.exp((-dtheta**2) / (2 * thetaSigma**2)).numpy() EnergyAll=np.zeros((rows,cols)) AnAll=np.zeros((rows,cols)) for o in range(1,norient+1): sumE_ThisOrient=np.zeros((rows,cols)) sumO_ThisOrient=np.zeros((rows,cols)) sumAn_ThisOrient=np.zeros((rows,cols)) Energy=np.zeros((rows,cols)) for s in range(1,nscale+1): filter = logGabor[s-1] * spread[o-1] ifftFilt=(tf.math.real(tf.signal.ifft2d(filter))*math.sqrt(rows*cols)).numpy() ifftFilterArray[s-1]=ifftFilt EO[s-1][o-1]=tf.signal.ifft2d(imagefft*filter).numpy() An=tf.math.abs(EO[s-1][o-1]).numpy() sumAn_ThisOrient = sumAn_ThisOrient + An sumE_ThisOrient = sumE_ThisOrient + tf.math.real(EO[s-1][o-1]).numpy() sumO_ThisOrient = sumO_ThisOrient + tf.math.imag(EO[s-1][o-1]).numpy() if s==1: EM_n=np.sum(filter**2) maxAn = An else: maxAn =tf.math.maximum(maxAn, An).numpy() XEnergy=tf.math.sqrt(sumE_ThisOrient**2+sumO_ThisOrient**2).numpy()+epsilon MeanE = sumE_ThisOrient / XEnergy MeanO = sumO_ThisOrient / XEnergy for s in range(1,nscale+1): E=tf.math.real(EO[s-1][o-1]).numpy() O=tf.math.imag(EO[s-1][o-1]).numpy() Energy = Energy + E*MeanE + O*MeanO - tf.math.abs(E*MeanO - O*MeanE).numpy() medianE2n = np.median(np.reshape((tf.math.abs(EO[0][o-1]).numpy())**2,(rows*cols,))) meanE2n = -medianE2n/math.log(0.5) estMeanE2n[o-1] = meanE2n noisePower = meanE2n/EM_n EstSumAn2=np.zeros((rows,cols)) for s in range(1,nscale+1): EstSumAn2 = EstSumAn2 + ifftFilterArray[s-1]**2 EstSumAiAj=np.zeros((rows,cols)) for si in range(1,nscale): for sj in range(si+1,nscale+1): EstSumAiAj = EstSumAiAj + ifftFilterArray[si-1]*ifftFilterArray[sj-1] sumEstSumAn2=np.sum(EstSumAn2) sumEstSumAiAj=np.sum(EstSumAiAj) EstNoiseEnergy2 = 2*noisePower*sumEstSumAn2 + 4*noisePower*sumEstSumAiAj tau = math.sqrt(EstNoiseEnergy2/2) EstNoiseEnergy = tau*math.sqrt(math.pi/2) EstNoiseEnergySigma = math.sqrt( (2-math.pi/2)*tau**2 ) T = EstNoiseEnergy + k*EstNoiseEnergySigma T = T/1.7 Energy=Energy-T Energy=tf.math.maximum(Energy,tf.zeros((rows,cols))).numpy() EnergyAll = EnergyAll + Energy AnAll = AnAll + sumAn_ThisOrient ResultPC = EnergyAll / AnAll return ResultPC def FeatureSIM(imageRef, imageDis): imageRef=np.squeeze(imageRef) imageDis=np.squeeze(imageDis) img_shape=np.shape(imageRef) rows,cols=img_shape[:2] I1=np.ones((rows,cols)) I2=np.ones((rows,cols)) Q1=np.ones((rows,cols)) Q2=np.ones((rows,cols)) if len(img_shape)==3: Y1 = 0.299 * imageRef[:,:,0] + 0.587 * imageRef[:,:,1] + 0.114 * imageRef[:,:,2] Y2 = 0.299 * imageDis[:,:,0] + 0.587 * imageDis[:,:,1] + 0.114 * imageDis[:,:,2] I1 = 0.596 * imageRef[:,:,0] - 0.274 * imageRef[:,:,1] - 0.322 * imageRef[:,:,2] I2 = 0.596 * imageDis[:,:,0] - 0.274 * imageDis[:,:,1] - 0.322 * imageDis[:,:,2] Q1 = 0.211 * imageRef[:,:,0] - 0.523 * imageRef[:,:,1] + 0.312 * imageRef[:,:,2] Q2 = 0.211 * imageDis[:,:,0] - 0.523 * imageDis[:,:,1] + 0.312 * imageDis[:,:,2] else: Y1=imageRef Y2=imageDis Y1=Y1.astype(np.float32) Y2=Y2.astype(np.float32) I1=I1.astype(np.float32) I2=I2.astype(np.float32) Q1=Q1.astype(np.float32) Q2=Q2.astype(np.float32) minDimension = min(rows,cols) F = max(1,round(minDimension / 256)) aveKernel=np.ones((F,F))/(F*F) aveKernel=aveKernel.astype(np.float32) aveI1=tf.nn.conv2d(tf.reshape(I1,[1,rows,cols,1]),tf.reshape(aveKernel,[F,F,1,1]),strides=[1,1,1,1],padding='SAME').numpy() aveI2=tf.nn.conv2d(tf.reshape(I2,[1,rows,cols,1]),tf.reshape(aveKernel,[F,F,1,1]),strides=[1,1,1,1],padding='SAME').numpy() I1=aveI1[0,0:rows:F,0:cols:F,0] I2=aveI2[0,0:rows:F,0:cols:F,0] aveQ1=tf.nn.conv2d(tf.reshape(Q1,[1,rows,cols,1]),tf.reshape(aveKernel,[F,F,1,1]),strides=[1,1,1,1],padding='SAME').numpy() aveQ2=tf.nn.conv2d(tf.reshape(Q2,[1,rows,cols,1]),tf.reshape(aveKernel,[F,F,1,1]),strides=[1,1,1,1],padding='SAME').numpy() Q1=aveQ1[0,0:rows:F,0:cols:F,0] Q2=aveQ2[0,0:rows:F,0:cols:F,0] aveY1=tf.nn.conv2d(tf.reshape(Y1,[1,rows,cols,1]),tf.reshape(aveKernel,[F,F,1,1]),strides=[1,1,1,1],padding='SAME').numpy() aveY2=tf.nn.conv2d(tf.reshape(Y2,[1,rows,cols,1]),tf.reshape(aveKernel,[F,F,1,1]),strides=[1,1,1,1],padding='SAME').numpy() Y1=aveY1[0,0:rows:F,0:cols:F,0] Y2=aveY2[0,0:rows:F,0:cols:F,0] PC1 = phasecong2(Y1) PC2 = phasecong2(Y2) dx=np.array([[3,0,-3],[10,0,-10],[3,0,-3]],np.float32)/16 dy=np.array([[3,10,3],[0,0,0],[-3,-10,-3]],np.float32)/16 IxY1=tf.nn.conv2d(tf.reshape(Y1,[1,Y1.shape[0],Y1.shape[1],1]),tf.reshape(dx,[3,3,1,1]),strides=[1,1,1,1],padding='SAME').numpy() IyY1=tf.nn.conv2d(tf.reshape(Y1,[1,Y1.shape[0],Y1.shape[1],1]),tf.reshape(dy,[3,3,1,1]),strides=[1,1,1,1],padding='SAME').numpy() gradientMap1=tf.math.sqrt(IxY1**2+IyY1**2).numpy()