Problem B (MCM) Industrial surface defect detection
Surface defects in metal or plastic products not only affect the appearance of the product, but may also cause serious damage to the performance or durability of the product. Automatic surface anomaly detection has become an interesting and promising research area with a very high and direct impact on the application areas of visual inspection [1]. The Collector Group provides a dataset of images of defective production items [2], we would like to use this dataset as an example to study a mathematical model for automatic detection
Product surface defects are captured through photos.
Doman Tabernik, Mattik UC, and Danijel Skoy built a model for detecting surface defects using deep learning [3], which is claimed to provide good recognition even with a small amount of training . However
, our problem at this point is slightly different; first, we want our model to be deployable on cheap handheld devices. Such devices have very limited storage space and computing power, so the model is very demanding in terms of computational effort and required storage space. Second, since this dataset does not contain all defect patterns, we expect the model to have relatively good generalization capabilities when encountering other defect types. You and your team are asked to build easy-to-use mathematical models to complete the following tasks.
Task:
- Determine whether surface defects appear in photographs and measure the computational effort and storage required for the model to do so;
- . Automatically mark locations or areas where surface defects occur and measure the computational effort, storage space, and marking accuracy required for the model.
- . Please clarify the generalization ability of your model, i. e. Why is your model still feasible if the types of defects you encounter are not entirely feasible?