图像质量评估：BRISQUE

Mat im = imread("image_scenery.jpg"); // read image

cvtColor(im, im, COLOR_BGR2GRAY); // convert to grayscale


im.convertTo(im, 1.0/255); // normalize and copy the read image to orig_bw
Mat mu(im.size(), CV_64FC1, 1);
GaussianBlur(im, mu, Size(7, 7), 1.166); // apply gaussian blu
Mat mu_sq = mu.mul(mu);


// compute sigma
Mat sigma = im.size(), CV_64FC1, 1);
sigma = im.mul(im); 
GaussianBlur(sigma, sigma, Size(7, 7), 1.166); // apply gaussian blu
subtract(sigma, mu_sq, sigma); // sigma = sigma - mu_sq
cv::pow(sigma, double(0.5), sigma); // sigma = sqrt(sigma)
add(sigma, Scalar(1.0/255), sigma); // to avoid DivideByZero Exception
Mat structdis(im.size(), CV_64FC1, 1);
subtract(im, mu, structdis); // structdis = im - mu
divide(structdis, sigma, structdis);

im = cv2.imread("image_scenery.jpg", 0) # read as gray scale
blurred = cv2.GaussianBlur(im, (7, 7), 1.166) # apply gaussian blur to the image
blurred_sq = blurred * blurred 
sigma = cv2.GaussianBlur(im * im, (7, 7), 1.166) 
sigma = (sigma - blurred_sq) ** 0.5
sigma = sigma + 1.0/255 # to make sure the denominator doesn't give DivideByZero Exception
structdis = (im - blurred)/sigma # final MSCN(i, j) image

// declare shifting indices array
int shifts[4][2] = {{0, 1}, {1, 0}, {1, 1}, {-1, 1}};


// calculate pair-wise products for every combination of shifting indices
for(int itr_shift = 1; itr_shift <= 4; itr_shift++) 
{
    int* reqshift = shifts[itr_shift - 1]; // the required shift index
    // declare shifted image
    Mat shifted_structdis(imdist_scaled.size(), CV_64F, 1);
    // BwImage is a helper class to create a copy of the image and create helper functions to access it's pixel values
    BwImage OrigArr(structdis);
    BwImage ShiftArr(shifted_structdis); 
    // calculate pair-wise component for the given orientation
    for(int i = 0; i < structdis.rows; i++)
    {
        for(int j = 0; j < structdis.cols; j++) { if(i + reqshift[0] >= 0 && i + reqshift[0] < structdis.rows && j + reqshift[1] >= 0 && j + reqshift[1] < structdis.cols)
            {
                ShiftArr[i][j] = OrigArr[i + reqshift[0]][j + reqshift[1]];
            }f
            else
            {
                ShiftArr[i][j] = 0;
            }
        }   
    }
    Mat shifted_new_structdis;
    shifted_new_structdis = ShiftArr.equate(shifted_new_structdis);
    // find the pairwise product
    multiply(structdis, shifted_new_structdis, shifted_new_structdis);
}

# indices to calculate pair-wise products (H, V, D1, D2)
shifts = [[0,1], [1,0], [1,1], [-1,1]]
# calculate pairwise components in each orientation
for itr_shift in range(1, len(shifts) + 1):
    OrigArr = structdis
    reqshift = shifts[itr_shift-1] # shifting index


    for i in range(structdis.shape[0]):
        for j in range(structdis.shape[1]):
            if(i + reqshift[0] >= 0 and i + reqshift[0] < structdis.shape[0]  and j + reqshift[1] >= 0 and j  + reqshift[1] < structdis.shape[1]):
               ShiftArr[i, j] = OrigArr[i + reqshift[0], j + reqshift[1]]
            else:
               ShiftArr[i, j] = 0

# create affine matrix (to shift the image)
M = np.float32([[1, 0, reqshift[1]], [0, 1, reqshift[0]]])
ShiftArr = cv2.warpAffine(OrigArr, M, (structdis.shape[1], structdis.shape[0])

// make a svm_node object and push values of feature vectors into it
struct svm_node x[37];


// features is a rescaled vector to (-1, 1)
for(int i =; i < 36; ++i) {
  x[i].value = features[i];
  x[i].index = i + 1; // svm_node indexes start from 1
}


// load the model from modelfile - allmodel
string model = "allmodel";
model = svm_load_model(modelfile.c_str())


// get number of classes from the model
int nr_class = svm_get_nr_class(model);
double *prob_estimates = (double *) malloc(nr_class*sizeof(double));


// predict the quality score using svm_predict_probability function
qualityscore = svm_predict_probability(model, x, prob_estimates);

# load the model from allmodel file
model = svmutil.svm_load_model("allmodel")
# create svm node array from features list
x, idx = gen_svm_nodearray(x[1:], isKernel=(model.param.kernel_type == PRECOMPUTED))
nr_classifier = 1 # fixed for svm type as EPSILON_SVR (regression)
prob_estimates = (c_double * nr_classifier)()


# predict quality score of an image using libsvm module 
qualityscore = svmutil.libsvm.svm_predict_probability(model, x, dec_values)

# Python 2.7 
python2 brisquequality.py <image_path>
# Python 3.x
python3 brisquequailty.py <image_path>
</image_path></image_path>