def psnr_metric(y_true, y_pred):
return psnr(y_true, y_pred, y_range)
def psnr(self, y_true, y_pred):
return tf.reduce_mean(tfimg.psnr(y_true, y_pred, max_val=self.max_val))
def psnr_pred(y_true, y_pred):
return image.psnr(y_true, y_pred, max_val=1.0)
img_outputs = utils.convert_to_image_domain(freq_outputs)
else:
img_outputs = model.predict(img_inputs)
freq_outputs = utils.convert_to_frequency_domain(img_outputs)
mag_outputs = np.expand_dims(np.abs(utils.join_reim(img_outputs)), -1)
ssim = image.ssim(tf.convert_to_tensor(mag_outputs, dtype=tf.float32),
tf.convert_to_tensor(mag_label, dtype=tf.float32),
np.max(mag_label),
filter_size=11,
filter_sigma=1.5,
k1=0.01,
k2=0.03).numpy()
psnr = image.psnr(tf.convert_to_tensor(mag_outputs, dtype=tf.float32),
tf.convert_to_tensor(mag_label, dtype=tf.float32),
np.max(mag_label)).numpy()
output_file = os.path.join(model_dir, exp_str + '_test_output.h5')
if os.path.exists(output_file):
os.remove(output_file)
file = h5py.File(output_file, "w")
dataset = file.create_dataset("img_outputs",
np.shape(img_outputs),
data=img_outputs)
dataset = file.create_dataset("freq_outputs",
np.shape(freq_outputs),
data=freq_outputs)
dataset = file.create_dataset("mag_outputs",
np.shape(mag_outputs),
def psnr(hr, sr):
return tfimg.psnr(hr, sr, 1)
data_path = path + '\\data\\'
log_file = path + '\\log_total.txt' # 储存计算psnr值的位置
n = 100 # 压缩100张bmp图片
lf = open(log_file, 'w')
for i in range(n):
bmp_file = data_path + str(i + 1) + '.bmp' # 第i+1张bmp图片
bmp_img = cv2.imread(bmp_file)
lf.write('{}.jpg\n'.format(i + 1))
for qf in range(1, 101): # 对1-100范围内的qf计算bpp和psnr的值
lf.write('qf: {}\n'.format(qf))
buffer = cv2.imencode('.jpg', bmp_img,
[int(cv2.IMWRITE_JPEG_QUALITY), qf])[1]
bpp_val = (buffer.size * 8) / (1280 * 720) # 该图片的码率bpp
jpg_img = cv2.imdecode(buffer, cv2.IMREAD_COLOR)
psnr_val = image.psnr(bmp_img, jpg_img, max_val=255) # 该图片的psnr
psnr_bpp_val = psnr_val / bpp_val # 遗留问题，本来打算作为其它用途，但是好像并没有用到，避免删改出现不必要的bug
lf.write('psnr: {}\n'.format(psnr_val))
lf.write('bpp: {}\n'.format(bpp_val))
lf.write('psnr_bpp: {}\n'.format(psnr_bpp_val))
lf.write('\n\n')
lf.close()
def PSNR(o_img, r_img_org):
width, height = 1280, 720
r_img = cv2.resize(r_img_org, (width, height), cv2.INTER_CUBIC)
return image.psnr(o_img, r_img, max_val=255).numpy()
def mean_psnr(gt_photos, outputs):
return np.mean(image.psnr(gt_photos, outputs, 1))
def PSNR(y_true, y_pred):
return tfimage.psnr(y_true, y_pred, max_val=1.0)