Label-Free Deep Learning with Feature Adaptation for Crop Anomaly Detection on Small Datasets

Abstract

Efficient crop health monitoring is crucial for global food security. Supervised deep learning approaches are often impractical due to the scarcity of large, labeled datasets. To address this limitation, this study adapts EfficientAD, an unsupervised, label-free anomaly detection framework originally designed for industrial inspection, for agricultural imagery on small datasets. The method utilizes a Patch Description Network (PDN) for localized feature extraction, a student network for local anomalies, and an autoencoder for global structural constraints. Benchmarked against AnoGAN, Pix2Pix, InTra, and Teacher–Student models, the framework demonstrated superior performance on the MVTec AD, PlantVillage, Coffee Leaf, and a custom real-world Sweet Potato dataset. The model achieved perfect area under the receiver operating characteristic curve (AUROC) scores of up to 100% in categories like “Pongamia”, “Potato”, and “Coffee Leaf”. While image-level classification was exceptionally robust, pixel-level localization (AUPRO) proved sensitive to complex agricultural backgrounds. To overcome this, a background interference analysis was conducted using Background Removed (BGRM) and out-of-distribution Background Replaced-Green (BGRP-G) strategies on the custom dataset. Notably, the BGRP-G strategy remarkably improved the image-level AUROC from 88.9% to 99.5% and substantially boosted the pixel-level AUPRO from 47.1% to 61.9%, successfully preserving the boundary integrity of severe structural defects. Achieving millisecond-level latency without complex data augmentation, this adapted label-free framework offers a versatile, highly efficient solution for real-time crop health diagnostics on resource-constrained Edge AI devices.

Plant phenotyping relevance

農作物画像から異常・健康状態を抽出する深層学習手法を開発し、複数データセットで性能比較・検証しているため、植物フェノタイピング手法が中心である。

Code and data availability

The supplied blocks describe the paper's methods and datasets but contain no authors' public code repository, no public release of the custom Sweet Potato dataset, and no trained model checkpoints. The datasets used (MVTec AD, PlantVillage, Plant Leaves, Coffee Leaf Disease) are cited prior public resources, not paper-

No evidence-backed public reproduction asset is currently recorded.