DeepBlue Technology's Robot Vision Algorithm Team Wins Kaggle Paddy Disease Classification Championship
Recently, the final results of the Paddy Disease Classification competition held on the Kaggle platform were announced, and the robot vision algorithm team formed by DeepBlue Technology won the championship. This is the second time that DeepBlue Technology has won the top AI competition in the world in just one month, following its victory in the "Conditional Hierarchical Multi-Answer Question-Answering" evaluation task competition at the 16th National Knowledge Graph and Semantic Computing Conference (CCKS 2022) on August 25 this year.
As we all know, computer vision is currently the most active research field in the academic AI community. International competitions in related fields have attracted industry technology giants such as Huawei, BAT, University of Science and Technology of China, Google, Stanford, and top universities around the world. The competition results are recognized by the industry as a basis for evaluating the technical and academic levels of enterprises or academic institutions.
Kaggle, as the world's largest data science community and data science competition platform, attracted a total of 657 participating teams from around the world in this competition. The participating teams include PSI, an active global ranking No. 3 data scientist from Austria on the Kaggle platform, Mathurin Aché, a No. 33 data scientist from France, and Jeremy Howard, a former No. 1 data scientist from Australia, and many other top global data analysis experts.
This competition focuses on the application of deep learning technology in the field of large-scale fine-grained plant disease analysis. The PaddyDisease dataset involved in the competition contains 10 categories of rice diseases, with small data volume, fine granularity, imbalanced category data, and high classification difficulty.
DeepBlue Technology's success in this competition is mainly attributed to the team's adoption of a new strategy in the selection and training of the classification model. In terms of the selection of the classification model's backbone network, in addition to using the classic EfficientNet series, the team also employed new algorithms such as Swin-transformer and Convnext, which have been developed recently. They used the K-folds strategy, specifically the method of 5-fold cross-validation, to partition the data during training. In terms of model validation, they used techniques such as model fusion and test-time augmentation (TTA). For data augmentation, they employed the auto_augment method to enhance the diversity of the data and improve the model's generalization ability.
In the final competition, DeepBlue Technology utilized model fusion with new algorithms, models, and methods, achieving a final practical result of 0.98963 points. They stood out from 657 participating teams and won the first place in the competition.
Based on the technologies obtained in this competition, the company can not only apply them to various classification needs in robot visual perception but also has a wide range of practical applications in fine-grained classification scenarios.
DeepBlue Technology has won more than 40 championships in top global artificial intelligence and computer events such as ICCV, CVPR, and ECCV. Winning the championship in the Kaggle platform competition, known as the "arena of technical warriors," not only demonstrates DeepBlue Technology's outstanding academic achievements and profound technical expertise in the field of computer vision but also showcases the company's relentless pursuit of artificial intelligence technological innovation.