In the context of textile fabric, rare anomaly can occurs, hence compromizing the quality of the tissues. In order to avoid that in some scenario, it is crucial to detect the defect.
Fabric defect detection is a significant phase of quality control in textile industry. Manual defect inspection lacks the accuracy and the labor cost is high. So the goal of this project is to develop a system to detect defects in fabrics with Image Processing techniques and Neural Networks.
The dataset is 4 GB and has 48000 distinct records. The size of the images in the dataset is either 32x32 or 64x64. The data includes records that belong 6 different classes of defects -
- good - If the fabric is good and has no defect
- color - There is a defect in the color of the fabric
- cut - There is a cut in the fabric
- hole - There is a hole in the fabric
- thread - There is some thread coming off
- metal contamination - Fabric has some metal contamination
The dataset inlcudes images the original images and images created after rotations. The data for 8 different rotations in - 0, 20, 40, 60, 80, 100, 120, 140 is included.Given an image size, a train and test dataset are available with randomly generated patches. Source images from the train and test are non-overlapping
We tried to train models with three different architectures.
- CNN
- ResNet18
- ResNet34
transforms.Compose([
transforms.ToTensor(), transforms.Normalize([0.3541], [0.1352]), transforms.RandomErasing(0.3, value=0)
])
| Models | Accuracy | Train Loss | Validation Loss |
|---|---|---|---|
| CNN | 0.82 | 1.303 | 1.333 |
| ResNet18 | 0.98 | 1.286 | 26.913 |
| ResNet34 | 0.95 | 1.333 | 21.588 |
- PREPROCESSING STEPS: Add more preprocessing steps like rotating at angles other than in dataset
- DIFFERENT MODELS : Try more complex neural network models like VQVAE
- ADD MORE DATA : Try to get more data to train the models better
ResNet18
Classification Task: Keepíng all the data
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [1, 64, 16, 16] 3,136
BatchNorm2d-2 [1, 64, 16, 16] 128
ReLU-3 [1, 64, 16, 16] 0
MaxPool2d-4 [1, 64, 8, 8] 0
Conv2dAuto-5 [1, 64, 8, 8] 36,864
BatchNorm2d-6 [1, 64, 8, 8] 128
ReLU-7 [1, 64, 8, 8] 0
Conv2dAuto-8 [1, 64, 8, 8] 36,864
BatchNorm2d-9 [1, 64, 8, 8] 128
ResNetBasicBlock-10 [1, 64, 8, 8] 0
Conv2dAuto-11 [1, 64, 8, 8] 36,864
BatchNorm2d-12 [1, 64, 8, 8] 128
ReLU-13 [1, 64, 8, 8] 0
Conv2dAuto-14 [1, 64, 8, 8] 36,864
BatchNorm2d-15 [1, 64, 8, 8] 128
ResNetBasicBlock-16 [1, 64, 8, 8] 0
ResNetLayer-17 [1, 64, 8, 8] 0
Conv2d-18 [1, 128, 4, 4] 8,192
BatchNorm2d-19 [1, 128, 4, 4] 256
Conv2dAuto-20 [1, 128, 4, 4] 73,728
BatchNorm2d-21 [1, 128, 4, 4] 256
ReLU-22 [1, 128, 4, 4] 0
Conv2dAuto-23 [1, 128, 4, 4] 147,456
BatchNorm2d-24 [1, 128, 4, 4] 256
ResNetBasicBlock-25 [1, 128, 4, 4] 0
Conv2dAuto-26 [1, 128, 4, 4] 147,456
BatchNorm2d-27 [1, 128, 4, 4] 256
ReLU-28 [1, 128, 4, 4] 0
Conv2dAuto-29 [1, 128, 4, 4] 147,456
BatchNorm2d-30 [1, 128, 4, 4] 256
ResNetBasicBlock-31 [1, 128, 4, 4] 0
ResNetLayer-32 [1, 128, 4, 4] 0
Conv2d-33 [1, 256, 2, 2] 32,768
BatchNorm2d-34 [1, 256, 2, 2] 512
Conv2dAuto-35 [1, 256, 2, 2] 294,912
BatchNorm2d-36 [1, 256, 2, 2] 512
ReLU-37 [1, 256, 2, 2] 0
Conv2dAuto-38 [1, 256, 2, 2] 589,824
BatchNorm2d-39 [1, 256, 2, 2] 512
ResNetBasicBlock-40 [1, 256, 2, 2] 0
Conv2dAuto-41 [1, 256, 2, 2] 589,824
BatchNorm2d-42 [1, 256, 2, 2] 512
ReLU-43 [1, 256, 2, 2] 0
Conv2dAuto-44 [1, 256, 2, 2] 589,824
BatchNorm2d-45 [1, 256, 2, 2] 512
ResNetBasicBlock-46 [1, 256, 2, 2] 0
ResNetLayer-47 [1, 256, 2, 2] 0
Conv2d-48 [1, 512, 1, 1] 131,072
BatchNorm2d-49 [1, 512, 1, 1] 1,024
Conv2dAuto-50 [1, 512, 1, 1] 1,179,648
BatchNorm2d-51 [1, 512, 1, 1] 1,024
ReLU-52 [1, 512, 1, 1] 0
Conv2dAuto-53 [1, 512, 1, 1] 2,359,296
BatchNorm2d-54 [1, 512, 1, 1] 1,024
ResNetBasicBlock-55 [1, 512, 1, 1] 0
Conv2dAuto-56 [1, 512, 1, 1] 2,359,296
BatchNorm2d-57 [1, 512, 1, 1] 1,024
ReLU-58 [1, 512, 1, 1] 0
Conv2dAuto-59 [1, 512, 1, 1] 2,359,296
BatchNorm2d-60 [1, 512, 1, 1] 1,024
ResNetBasicBlock-61 [1, 512, 1, 1] 0
ResNetLayer-62 [1, 512, 1, 1] 0
ResNetEncoder-63 [1, 512, 1, 1] 0
AdaptiveAvgPool2d-64 [1, 512, 1, 1] 0
Linear-65 [1, 6] 3,078
ResnetDecoder-66 [1, 6] 0
================================================================
Total params: 11,173,318
Trainable params: 11,173,318
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.00
Forward/backward pass size (MB): 1.23
Params size (MB): 42.62
Estimated Total Size (MB): 43.86
----------------------------------------------------------------
Total_params 11173318
Trainable_params 11173318
CNN
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [1, 32, 30, 30] 320
BatchNorm2d-2 [1, 32, 30, 30] 64
ReLU-3 [1, 32, 30, 30] 0
AvgPool2d-4 [1, 32, 15, 15] 0
Conv2d-5 [1, 64, 13, 13] 18,496
BatchNorm2d-6 [1, 64, 13, 13] 128
ReLU-7 [1, 64, 13, 13] 0
AvgPool2d-8 [1, 64, 6, 6] 0
Conv2d-9 [1, 32, 6, 6] 2,080
AdaptiveAvgPool2d-10 [1, 32, 1, 1] 0
Linear-11 [1, 6] 198
================================================================
Total params: 21,286
Trainable params: 21,286
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.00
Forward/backward pass size (MB): 0.99
Params size (MB): 0.08
Estimated Total Size (MB): 1.07
----------------------------------------------------------------
Total_params 21286
Trainable_params 21286
Contributors: Emre Okcular, Hashneet Kaur