- To train the customized deep neural network on the device cover images
- To develop the image classification application with the trained model to infer the real-time status of webcam images for automated optical insepction (AOI) applications
- Convert the color space of the image from BGR to RGB
- Resize the image to 224 x 224
- Subtract the per-channel mean of the imagenet dataset (RGB:[123.68, 116.779, 103.939]) from the resized image
- Dataset: 600 images (PASS), 600 images (FAIL)
Experiment-I: Transfer-learning the pre-trained VGG model on images captured at limited distances from the camera
- Refer to this link for the training platform setup
- Model Training: Execution Time: 2.45 m, Epochs: 57, Trained Model: best_model_14-01-2021_10:56:44.h5
- Save the model after each epoch in the ModelCheckPoint and set patience = 30 in the EarlyStopping
PASS
FAIL
Failed to achieve higher accuracy on testing images which are captured at different angles and varying distances from the camera
Created the training and validation datasets with the images covering the entire device cover
Sample Images
PASS
FAIL
Experiment-II: Transfer-learning the pre-trained VGG model on images captured at different angles and varying distances from the camera
- Dataset: PASS: 1000 images per class
- Resize the image to 224 x 224
- Subtract the RGB values of each pixel from the mean RGB of the imagenet dataset. Set batch size to 32
- Freeze all the layers in the VGG network. FC network: 2 x 512-node hidden layer+2-node output layer with softmax classifier. Train the network on the target dataset.
- Epochs: 42, Time: 2.8 m, Trained Model: best_model_30-03-2021_13:29:22.h5
Better than Experiment-I, but still failed on testing images captured at different angles which are not included in the training dataset
- Dataset: PASS: 1000 images per class
- Resize the image to 224 x 224
- Subtract the RGB values of each pixel from the mean RGB of the imagenet dataset. Set batch size to 32
- Unfreeze the last two blocks of conv layers in the VGG network. FC network: 2 x 512-node hidden layer+2-node output layer with softmax classifier. Train the network on the target dataset
- Epochs: 30, Time: 3.4 m, Trained Model: best_model_30-03-2021_11:58:56.h5
PASS
FAIL
Achieves the best performance on testing images captured at different angles
Evaluate the performance of models trained in Experiment-I and Experiment-II
