Demo the welding defect detection¶
The model is hosted in AlfredWJLu/WeldingDefect-V1 @ ModelScope, and trained on a toy dataset.
In [1]:
import warnings
warnings.filterwarnings("ignore", category=UserWarning)
import requests
import base64, json
import numpy as np
from matplotlib import pyplot as plt
import cv2
from det.exp import *
from det.data.data_augment import *
from det.utils import *
from utils import *
from dotenv import load_dotenv, find_dotenv
from run import appMissTags, Predictor, Exp, prepare_det, calculate_overlap
import os, traceback, glob
detector = prepare_det()
load_dotenv(find_dotenv())
Out[1]:
In [ ]:
img_loc = 'YOU IMG LOC'
img_id = os.path.basename(img_loc).split('.')[0]
img_np = cv2.imread(img_loc,flags=cv2.IMREAD_GRAYSCALE)
img_np.shape
Out[ ]:
In [3]:
chunk_size = 640
eval_img_loc = str(os.getenv("EVAL_IMG_LOC"))
eval_img_loc = os.path.join(eval_img_loc, img_id)
os.makedirs(os.path.join(eval_img_loc, img_id),exist_ok=True)
'''make chunks'''
img_np_x1, img_np_y1, img_np_x2, img_np_y2 = 0, 0, img_np.shape[1], img_np.shape[0]
img_width, img_height = img_np.shape[1], img_np.shape[0]
print('Received a dicom image shape = {:}(W) x {:}(H)'.format(img_width, img_height))
width_overlap, height_overlap = calculate_overlap(img_width, img_height, chunk_size)
print("W overlap = {:}, H overlap = {:}".format(width_overlap, height_overlap))
num_chunks_width = (img_width + chunk_size - 1) // chunk_size
num_chunks_height = (img_height + chunk_size - 1) // chunk_size
for i in range(num_chunks_height):
for j in range(num_chunks_width):
# 计算当前 chunk 的左上角和右下角坐标
x_start = j * (chunk_size - width_overlap)
y_start = i * (chunk_size - height_overlap)
x_end = x_start + chunk_size
y_end = y_start + chunk_size
# 确保不超过图片边界
x_start = max(0, x_start)
y_start = max(0, y_start)
x_end = min(img_width, x_end)
y_end = min(img_height, y_end)
img_slide = img_np[y_start:y_end, x_start:x_end]
print(np.amin(img_slide), np.amax(img_slide))
print("Chunk slide H {:04}-{:04}, W {:04}-{:04}".format(y_start, y_end, x_start, x_end))
img_slide_name = 'Chunk.H{:04}-H{:04}-W{:04}-W{:04}.jpg'.format(\
y_start, y_end, x_start, x_end)
img_slide_name = os.path.join(eval_img_loc, img_slide_name)
cv2.imwrite(img_slide_name, np.array(img_slide).astype(np.uint8))
'''make detection on these chunks'''
files = glob.glob(os.path.join(eval_img_loc, "Chunk*.jpg"))
result_image = np.zeros((img_np.shape[0],img_np.shape[1],3))
is_ok = True
for image_name in files:
try:
print('Infer {:}'.format(image_name))
outputs, img_info = detector.inference(image_name)
image_basename = os.path.basename(image_name)
image_loc_part = str(image_basename.split('.')[1]).split('-')
image_loc_int = [int(x[1:]) for x in image_loc_part]
print("Patching back slide H {:}-{:}, W {:}-{:}".format(image_loc_int[0], image_loc_int[1], image_loc_int[2], image_loc_int[3]))
patch_rlt, objs_in_img = detector.visual(outputs[0], img_info, detector.confthre)
result_image[image_loc_int[0]:image_loc_int[1],image_loc_int[2]:image_loc_int[3]] = patch_rlt
if len(objs_in_img) > 0:
is_ok = False
except Exception as e:
err_msg = "Failed to Infer {:}".format(image_name)
err_tb = traceback.format_exc()
print(err_msg)
print(err_tb)
In [4]:
from matplotlib import pyplot as plt
plt.imshow(result_image.astype(np.uint8))
Out[4]:
