Open Set Self and Across Domain Adaptation for Tomato Disease Recognition With Deep Learning Techniques

• Published in: Frontiers in plant science Vol. 12; p. 758027
• Main Authors: Fuentes, Alvaro, Yoon, Sook, Kim, Taehyun, Park, Dong Sun
• Format: Journal Article
• Language: English
• Published: Switzerland Frontiers Media S.A 10-12-2021
• Subjects: domain adaptation; new environments; open-set recognition; plant diseases; Plant Science; unknown data; plant diseases; domain adaptation; new environments; open-set recognition; unknown data
• ISSN: 1664-462X; 1664-462X
• DOI: 10.3389/fpls.2021.758027

Recent advances in automatic recognition systems based on deep learning technology have shown the potential to provide environmental-friendly plant disease monitoring. These systems are able to reliably distinguish plant anomalies under varying environmental conditions as the basis for plant intervention using methods such as classification or detection. However, they often show a performance decay when applied under new field conditions and unseen data. Therefore, in this article, we propose an approach based on the concept of open-set domain adaptation to the task of plant disease recognition to allow existing systems to operate in new environments with unseen conditions and farms. Our system specifically copes diagnosis as an open set learning problem, and mainly operates in the target domain by exploiting a precise estimation of unknown data while maintaining the performance of the known classes. The main framework consists of two modules based on deep learning that perform bounding box detection and open set self and across domain adaptation. The detector is built based on our previous filter bank architecture for plant diseases recognition and enforces domain adaptation from the source to the target domain, by constraining data to be classified as one of the target classes or labeled as unknown otherwise. We perform an extensive evaluation on our tomato plant diseases dataset with three different domain farms, which indicates that our approach can efficiently cope with changes of new field environments during field-testing and observe consistent gains from explicit modeling of unseen data.