Improving robustness by balancing edge, texture, and background features | |
| Author | Jirasak Buranathawornsom |
| Call Number | AIT RSPR no.DSAI-24-01 |
| Subject(s) | Neural networks (Computer science) Image processing |
| Note | A research study submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Data Science and Artificial Intelligence |
| Publisher | Asian Institute of Technology |
| Abstract | Studies have shown that while the current image recognition models perform well on standard benchmarks, their performance drop drastically when evaluated on out-of-distribution samples. Upon investigation, It has been shown that modern neural networks have a tendency to learn and make a prediction based on texture of the image rather than the high level shape, which contradicts humans’ behaviour. In this work, we aim to alleviate such shortcoming by propose a fine-tuning method that separately learn edge, texture and background features. Experiments show that by decouple foreground into edge features and texture features in the learning process, our method outperform the previous method that only learn on foreground object. Further analysis also confirm the importance of edge features, consistently achieving better performance compared to emphasizing tex ture features across robustness benchmark datasets. |
| Year | 2024 |
| Type | Research Study Project Report (RSPR) |
| School | School of Engineering and Technology |
| Department | Department of Information and Communications Technologies (DICT) |
| Academic Program/FoS | Data Science and Artificial Intelligence (DSAI) |
| Chairperson(s) | Chaklam Silpasuwanchai; |
| Examination Committee(s) | Mongkol Ekpanyapong |
| Scholarship Donor(s) | Royal Thai Government Fellowship; |
| Degree | Research Studies Project Report (M. Eng.) - Asian Institute of Technology, 2024 |