Severstal Defect Detection Challenge
This repository contains the notebook of the model and EDA for the Kaggle competition . The main goal of this challenge involved predicting the class of defects on a steel plate as well as its location for improving manufacturing quality. Dataset used for this problem comprised two parts, training and testing. The training data contained 12600 different images along with the encoded pixels of defect location, and the testing data contained 5506 images. In the data preprocessing part, converted the RLE (Run Length Encoding) to a mask where it masked the defective region from the remaining image. For this task, I trained a convolutional model with U-net and dilated convolution layers. Apart from this, I used the dice coefficient metrics for evaluating the model, and it attained an accuracy of 76%.
In the process of building the model, I read and implemented several models like U-net https://arxiv.org/abs/1505.04597, seg-net https://arxiv.org/pdf/1511.00561.pdf, squeeze-segnet https://arxiv.org/abs/1711.05491. Apart from this, I incorporated the hyper-parameter tuning, batch normalization, and optimization techniques to improve the model score.