def _USPS(self):
def resize_and_scale(img, size, scale):
img = skresize(img, size)
return 1 - (np.array(img, "float32") / scale)
# if os.path.isfile(self.data_path+'/USPS'+set+'.pt'):
sz = (28, 28)
imgs_usps = []
lbl_usps = []
if 'USPdata.zip' not in os.listdir(self.data_path):
urllib.request.urlretrieve(
'https://github.com/darshanbagul/USPS_Digit_Classification/raw/master/USPSdata/USPSdata.zip',
self.data_path + '/USPSdata.zip')
zip_ref = zipfile.ZipFile(self.data_path + '/USPSdata.zip', 'r')
zip_ref.extractall(self.data_path)
zip_ref.close()
if self.set == 'train' or self.set == 'validation':
for i in range(10):
label_data = self.data_path + '/Numerals/' + str(i) + '/'
img_list = os.listdir(label_data)
for name in img_list:
if '.png' in name:
img = skimread(label_data + name)
img = sk_rgb2gray(img)
resized_img = resize_and_scale(img, sz, 255)
imgs_usps.append(resized_img.flatten())
lbl_usps.append(i)
elif self.set == 'test':
test_path = self.data_path + '/Test/'
strt = 1
for lbl, cntr in enumerate(range(151, 1651, 150)):
for i in range(strt, cntr):
i = format(i, '04d')
img = skimread(
os.path.join(test_path, 'test_' + str(i) + '.png'))
img = sk_rgb2gray(img)
resized_img = resize_and_scale(img, sz, 255)
imgs_usps.append(resized_img.flatten())
lbl_usps.append(9 - lbl)
strt = cntr
# os.remove(self.data_path+'/USPSdata.zip')
shutil.rmtree(self.data_path + '/Numerals')
shutil.rmtree(self.data_path + '/Test')
imgs_usps, lbl_usps = np.asarray(imgs_usps).reshape(
-1, 28, 28), np.asarray(lbl_usps)
lbl_usps = torch.tensor(lbl_usps, dtype=torch.long)
imgs_usps = torch.tensor(imgs_usps)
# torch.save((imgs_usps,lbl_usps), open(self.data_path+'/USPS'+set+'.pt','wb'))
#
# else:
# imgs_usps, lbl_usps = torch.load(open(self.data_path+'/USPS'+set+'.pt','rb'))
return imgs_usps, lbl_usps
def imread(fname, factor=100):
"""Read possibly scaled version of image"""
img = skimread(fname)
if factor < 100:
img = imresize(img, [
int(img.shape[0] * factor / 100),
int(img.shape[1] * factor / 100)
],
order=3)
img = (img * 255).astype(np.uint8)
return img
def imread(
filepath: str,
return_type: Union[np.dtype, str] = np.uint8,
convert_to_tensor: bool = False,
) -> np.ndarray:
"""Read image data from file into the specified format.
This function wraps scikit-image imread functionality and applies required
conversions.
Args:
filepath: string containing path to the image data; .npy files are also
supported.
return_type: data format of the returned image, can be np.dtype or
a string. Supported formats are: float, uint8, uint16.
Default is uint8.
convert_to_tensor: flag to convert output image to torch.Tensor.
Defauls is False.
Returns:
Image in selected format."""
str_to_type = {
'float': np.float,
'uint8': np.uint8,
'uint16': np.uint16,
}
if hasattr(return_type, 'lower'):
return_type = return_type.lower()
if isinstance(return_type, str):
return_type = str_to_type[return_type]
# load file
if filepath.endswith('.npy'):
image = np.load(filepath)
else:
image = skimread(filepath)
# apply conversions
if not np.issubdtype(image.dtype, return_type):
if return_type is np.float:
image = convert_image_to_float(image)
else:
image = convert_image_to_int(image, image_type=return_type)
if convert_to_tensor:
image = ToTensor()(image)
return image
def imread(imgfile):
return th.from_numpy(skimread(imgfile))
def spot_scan(file):
try:
img_dir = os.path.dirname(file) # Determine directory of image file. Use to save outputs later.
##### LOAD IMAGE FILE AND CROP A D45 mm AREA FROM CENTER #####
img = skimread(file) # Read image file
# Determine image size for centered crop
height = img.shape[0]
width = img.shape[1]
center_y = int(height/2)
center_x = int(width/2)
img_center = (center_x, center_y)
mask_radius = 4252
# Create D45mm circular mask at center of image
mask = np.zeros((height,width), np.uint8)
cv2.circle(mask, img_center, mask_radius, (255, 255, 255), -1)
masked_img = cv2.bitwise_and(img, img, mask=mask)
# Threshold and find contours
thresh_mask = cv2.threshold(mask, 1, 255, cv2.THRESH_BINARY)[1]
mask_contours = cv2.findContours(thresh_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
x,y,w,h = cv2.boundingRect(mask_contours[0])
# Crop masked data
crop_img = masked_img[y:y+h,x:x+w]
#convert black edge pixels to white so that edge does not get captured by threshold
crop_img[np.where(crop_img==[0])] = 255
##### APPLY THRESHOLDING TO IMAGE THEN FIND AND RECORD MICROFEATURES #####
# Apply thresholding to the cropped image
thresh_img = cv2.threshold(crop_img, 145, 255, cv2.THRESH_BINARY_INV)[1]
# Create contours
(_,cnts,_) = cv2.findContours(thresh_img.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
cnt_data = {"Area":[], "Location":[]} # Initialize contour data
#determine if contour area is large enough to be interesting
for contour in cnts:
area = cv2.contourArea(contour)*0.000028
(x, y), radius = cv2.minEnclosingCircle(contour)
center = (int(x), int(y))
radius = int(radius)
cnt_data["Area"].append(area)
cnt_data["Location"].append(center)
if area < 0.002:
continue
else:
# If mf large enough draw a circle around the contour
cv2.circle(crop_img, center, radius, (0, 0, 255), 10)
# Create dataframe from cnt_data
cnt_data = pd.DataFrame(cnt_data)
# Resize and save image with circled microfeatures
out_img_file = img_dir + "/highlighted_spots.jpg" # Concatenate directory/filename for output image
resize_img = cv2.resize(crop_img, (800, 800))
cv2.imwrite(out_img_file, resize_img)
#cv2.imshow("original", resize_img)
#cv2.waitKey(0)
##### CREATE A HISTOGRAM OF MICROFEATURE SIZE #####
# create microfeature histogram
bins = [0.002, 0.0025, 0.003, 0.004, 0.006, 0.008, 0.012, 0.018, 0.025, 0.050, 0.1, 10]
bars = [0.0025, 0.003, 0.004, 0.006, 0.008, 0.012, 0.018, 0.025, 0.050, 0.1, 'More']
# Use hist method to sort data into bins
n, bins, patches = plt.hist(x=list(cnt_data["Area"]), bins=bins, color='#0504aa', alpha=0.7, rwidth=0.85)
bar_height = n.tolist() # Convert array of bin counts to list
y_pos = np.arange(len(bars)) # y_pos determines # of bars
plt.bar(y_pos, bar_height, align='center', alpha=0.5)
plt.xticks(y_pos, bars, rotation = 30)
plt.ylabel('Count')
plt.xlabel('Microfeature Size ($mm^2$)')
plt.title('Microfeature Distribution')
maxmf = max(list(cnt_data['Area'])) # Determine size of largest MF
text_y = max(bar_height) / 2 # Variable to place text at mid height of chart
plt.text(7, text_y , 'largest feature\n%s mm2' % str(round(maxmf,4)))
out_plt_file = img_dir + "/histogram.jpg"
plt.savefig(out_plt_file)
#plt.show()
##### CREATE A SUMMARY REPORT IN EXCEL #####
# Create summary statistics
nsmall = bar_height[0] + bar_height[1] + bar_height[2] + bar_height[3] + bar_height[4] # 0.002 to 0.008
nmed = bar_height[5] + bar_height[6] # 0.008 to 0.018
nlarge = bar_height[7] + bar_height[8] + bar_height[9] + bar_height[10] # 0.018 to 10
mf_coverage = sum(cnt_data["Area"].values)/(1590)*100
# Determine pass or fail
if nsmall < 380:
tsmall = "pass"
else:
tsmall = "fail"
if nmed < 250:
tmed = "pass"
else:
tmed = "fail"
if nlarge < 160:
tlarge = "pass"
else:
tlarge = "fail"
if maxmf < 0.1:
tmax = "pass"
else:
tmax = "fail"
# Set indexes and columns
index = ["Count 0.002 - 0.008 mm2", "Count 0.008 - 0.018 mm2", "Count 0.018 - 0.1 mm2", "Largest Spot", "Spot Coverage %"]
summary = {"Key Data":[nsmall, nmed, nlarge, round(maxmf,4), round(mf_coverage,4)], "Pass / Fail":[tsmall, tmed, tlarge, tmax, ""]}
summary = pd.DataFrame(data = summary, index = index)
# Send summary dataframe to excel then add a tab for cnt_data
out_xls_file = img_dir + "/spot_data.xlsx"
writer = pd.ExcelWriter(out_xls_file, engine = 'openpyxl')
summary.to_excel(writer, sheet_name = "Summary")
cnt_data.to_excel(writer, sheet_name = "Spot Data")
writer.save()
writer.close()
return 0, summary
except Exception as e:
return 1, ("scan error\n" + str(e))
def imread(filename : str) -> Image:
from skimage.io import imread as skimread
image = skimread(filename)
from .._tier0 import push
return push(image)
def imread(imgfile):
return skimread(imgfile)
def imread(image_path):
# Takes an image file (.png, .jpg etc) and returns a 2D np.array
return img_as_float(skimread(image_path))
