A two-stage computer vision platform for agriculture: automated data curation, perceptual deduplication, and anti-leakage audit (AgriVision), coupled with privacy-preserving collaborative model training across distributed cooperative nodes via Flower (AgriFed).
Agricultural diagnostic models require massive, diverse image datasets from multiple field organizations. However, agricultural cooperatives cannot share private visual data due to confidentiality and network constraints, and raw datasets often suffer from label noise, blurry captures, and cross-split data leakage.
Built an end-to-end two-phase solution: 1) AgriVision Khaos cleans, deduplicates via perceptual hashing, resolves label typos, and packages datasets under 2 GB for CKAN; 2) AgriFed Khaos trains vision backbones collaboratively across on-premise cooperative nodes using Flower, transferring only model weights to the central master aggregator.
Developed the autonomous curation CLI (`make run`), MongoDB metadata tracking, FiftyOne embeddings pipeline, and Flower federated simulation benchmarks, enabling multi-organization model training with zero raw data transfer.