Defect Detection in Metal Surfaces Using Computer Vision

Encoder models have shown remarkable success in various computer vision operations like object detection, image classification, and semantic segmentation. However, the results of one model showed that it was underfitting as it performed better on the validation set than on the training set. To gain insights into how the model was processing input images at different layers and which image features were being detected by different filters in the convolutional layers, an algorithm was used to visualize the activation maps for the output of convolutional layers in the model. To address the underfitting issue, the previous model was converted into a single feature-vector using global average pooling and used as a single layer in another deep learning model. As a result, the new model displayed a batch of images along with their corresponding segmentation masks and predicted segmentation masks generated by the encoder model. The output was a 3 × 1 image grid that included the original image, its ground truth segmentation mask, and the predicted segmentation mask. To optimize the model further, an algorithm was applied that trained an image encoder model on training data and evaluated it on validation data for a number of epochs. The algorithm then displayed a batch of samples with their ground truth and predicted segmentation masks. The results for this model were better than the previous model, as it was performing better with the train set as compared to the validation set. Overall, the usage of an encoder model and subsequent modifications allowed for improved performance in image segmentation tasks.