点击下方卡片,关注“机器视觉与AI深度学习”
视觉/图像重磅干货,第一时间送达!
导 读
本文主要介绍基于Segformer实现PCB缺陷检测 ,并给出步骤和代码。
背景介绍
PCB缺陷检测是电子制造的一个重要方面。利用Segformer等先进模型不仅可以提高准确性,还可以大大减少检测时间。传统方法涉及手动检查,无法扩展且容易出错。利用机器学习,特别是 Segformer模型,提供自动化且精确的解决方案。
实现步骤
下面是具体步骤:
【1】安装所需环境。首先,我们安装所需的库。其中,albumentations用于数据增强,transformers允许访问 Segformer等预训练模型,并xmltodict帮助解析数据集的XML注释。
pip install evaluate albumentations transformers accelerate xmltodict
【2】数据集。这个项目中使用的数据集由Roboflow提供。可以从下面链接获取:
https://universe.roboflow.com/diplom-qz7q6/defects-2q87r/dataset/16
该数据集分为测试文件夹和训练文件夹,由XML格式的图像及其相应注释组成。
# Create train and test sets
train_folder = "drive/..../train/images/"
test_folder = "drive/.../validation/images/"
train_img_paths = sorted([train_folder + f for f in os.listdir(train_folder) if f.endswith("jpg")])
test_img_paths = sorted([test_folder + f for f in os.listdir(test_folder) if f.endswith("jpg")])
train_xml_paths = [f[:-3] + "xml" for f in train_img_paths]
test_xml_paths = [f[:-3] + "xml" for f in test_img_paths]
train_ds = {"image_paths": train_img_paths, "xml_paths": train_xml_paths}
test_ds = {"image_paths": test_img_paths, "xml_paths": test_xml_paths}
【3】缺陷标注解析。对于每个缺陷标注信息:识别缺陷类型,提取缺陷的多边形形状,该多边形被绘制到初始化的蒙版上。最后,该函数弥合了XML标注信息和适合训练的格式之间的差距。给定 PCB图像及其相应的XML 注释,它会生成一个分割掩模,突出显示有缺陷的区域。掩模可以是适合训练模型的数字格式,也可以是用于人工检查的视觉格式。
def process_mask(img_path, xml_path, visualize=False):
img = cv2.imread(img_path)
num_dim = 3 if visualize else 1
mask = np.zeros((img.shape[0], img.shape[1], num_dim))
# Read xml content from the file
with open(xml_path, "r") as file:
xml_content = file.read()
data = xmltodict.parse(xml_content)
# If has defect mask
if "object" in data["annotation"]:
objects = data["annotation"]["object"]
# Single defects are annotated as a single dict, not a list
if not isinstance(objects, List):
objects = [objects]
for obj in objects:
defect_type = obj["name"]
polygon = obj["polygon"]
poly_keys = list(polygon.keys())
# Get number of (x, y) pairs - polygon coords
poly_keys = [int(k[1:]) for k in poly_keys]
num_poly_points = max(poly_keys)
# Parse ordered polygon coordinates
poly_coords = []
for i in range(1, num_poly_points+1):
poly_coords.append([
int(float(polygon[f"x{i}"])),
int(float(polygon[f"y{i}"]))
])
poly_coords = np.array(poly_coords)
# Draw defect segment on mask
fill_color = color_map[defect_type] if visualize else id_cat_map[defect_type]
mask = cv2.fillPoly(mask, pts=[poly_coords], color=fill_color)
#Optional
if visualize:
cv2.imwrite("output.jpg", mask)
mask = Image.open("output.jpg")
return mask
【4】探索性数据分析。在训练模型之前,最好先了解数据。在这里,我们检查缺陷类型的分布并在样本图像上可视化缺陷。
缺陷热力图显示了常见的缺陷位置,箱线图显示了缺陷尺寸的分布。
该函数旨在通过读取边界框详细信息来计算 XML 注释中存在的每个缺陷的大小。
def get_defect_sizes(xml_paths):
sizes = []
for xml_path in xml_paths:
with open(xml_path) as f:
data = xmltodict.parse(f.read())
objects = []
if 'object' in data['annotation']:
objects = data['annotation']['object']
if not isinstance(objects, list):
objects = [objects]
for obj in objects:
bndbox = obj['bndbox']
width = int(bndbox['xmax']) - int(bndbox['xmin'])
height = int(bndbox['ymax']) - int(bndbox['ymin'])
sizes.append(width * height)
return sizes
最后,群图重点关注缺陷尺寸在整个数据集中的分布和扩散。
【5】数据增强。该albumentations库用于人为扩展训练数据集,有助于提高模型的泛化能力。唯一指定的增强是水平翻转,它将以 50% 的概率水平翻转图像。
transform = A.Compose([
A.HorizontalFlip(p=0.5)
])
【6】图像预处理。将图像及其掩模预处理为适合Segformer模型的格式。
preprocessor = SegformerImageProcessor()
使用 OpenCV 加载图像。使用前面讨论的函数生成缺陷掩模process_mask。使用之前初始化的图像预处理图像及其掩模SegformerImageProcessor。此步骤将图像转换为张量格式,并确保它们具有适合 Segformer 模型的大小和标准化。返回预处理的图像和掩模张量。
class DefectSegmentationDataset(Dataset):
def __init__(self, dataset, mode):
self.image_paths = dataset["image_paths"]
self.xml_paths = dataset["xml_paths"]
def __len__(self):
return len(self.image_paths)
def __getitem__(self, idx):
# Read image
image = cv2.imread(self.image_paths[idx])
# Get mask
mask = process_mask(self.image_paths[idx], self.xml_paths[idx])
mask = mask.squeeze()
mask = Image.fromarray(mask.astype("uint8"), "L")
# Return preprocessed inputs
inputs = preprocessor(image, mask, ignore_index=None, return_tensors="pt")
inputs["pixel_values"] = inputs["pixel_values"].squeeze(0)
inputs["labels"] = inputs["labels"].squeeze(0)
return inputs
—THE END—
觉得有用,麻烦给个赞和在看