Explainable defect detection using convolutional neural networks: case study
Task
You are given a 400-image dataset, that contains images of good items (labeled as class ‘Good’) and items with a defect (labeled as class ‘Anomaly’).
Dataset is imbalanced — with more samples of good images than defective ones. Item in the image may be literally of any type and complexity — bottle, cable, pill, tile, leather, zipper, etc.
Your task is to build a model, that classifies images into ‘Good’ / ‘Anomaly’ classes and returns a bounding box for the defect if the image is classified as an ‘Anomaly’.
Even though this task may look simple, like a typical object detection task, there is an issue — we do not have labels for bounding boxes.
Model
Let’s take VGG16 pre-trained on ImageNet, and change its classification head — replace Flattening and Dense layers with Global Average Pooling and a single Dense layer.
We train the model as a typical 2-class classification model. The model outputs a 2-dimensional vector that contains probabilities for classes ‘Good’ and ‘Anomaly’.
Tài liệu tham khảo
Internet
Hết.
