APPLICATION OF LLM + ZERO-SHOT LARGE MODELS FOR FRUIT OBJECT DETECTION

A hierarchical prototype-graph with optimal-transport matching for few-shot rice disease recognition.

Accurate identification of rice diseases from field images is critical for crop health monitoring and sustainable agriculture, particularly in low-resource environments. However, most deep learning approaches depend on large-scale labeled datasets and pretrained backbones, limiting their applicability to rare or emerging diseases. In this work, we formulate a domain-specific prototype-based few-shot framework that avoids pretrained visual backbones and treats rice disease recognition as structured matching over a pathogen-aware class graph. The individual components, including wavelet-scattering features, optimal transport, semantic prototype fusion, and transductive refinement, are established techniques; the contribution lies in their coupled use within a disease-taxonomy-guided few-shot matching process. This design combines fixed visual descriptors, root-to-leaf prototype matching, class symptom descriptors, and confidence-gated refinement to support rice disease recognition under limited labeled data. We evaluate the model on two publicly available rice disease datasets-the Philippines Rice Diseases and Roboflow Rice-under 1-shot and 5-shot classification settings. In in-domain experiments, our approach achieves up to 95.8% accuracy and 94.9% macro-F1 on the Philippines dataset, consistently outperforming a diverse set of baselines including CNN-from-scratch, ResNet-18-from-scratch, Matching Networks, MAML, ProtoNet, RelationNet, SimpleShot, FEAT, and a flat optimal-transport variant. In cross-domain evaluation, the model demonstrates strong generalization capability, attaining up to 91.7% accuracy and 90.6% macro-F1 when transferring across datasets. An ablation study further confirms the consistent contribution of hierarchical structure, semantic fusion, and transductive refinement to performance gains. These results demonstrate that the proposed framework delivers highly accurate, robust, and data-efficient disease recognition, making it well-suited for real-world agricultural deployment under limited supervision.

ITR: Iterative Transductive Refinement with a Graph-Reliability Gate for Zero-Shot Agricultural CLIP Classification

Contrastive Language–Image Pre-training (CLIP) has become the dominant paradigm for zero-shot visual recognition, classifying images directly from textual class descriptions with no task-specific labelled data. This label-free ability is a natural fit for agricultural plant-disease and weed recognition, where expert annotation is costly and the space of crops, diseases, and weeds continually expands with new classes, so a classifier that labels an unlabelled batch of field images without re-training is attractive. Off-the-shelf zero-shot CLIP, however, falls well short on this fine-grained domain: agricultural imagery is under-represented in pre-training, classes differ only in subtle textural cues, and the posterior skews systematically toward a few generic labels. We present ITR (Iterative Transductive Refinement), a parameter-free, training-free method that refines the frozen-CLIP zero-shot posterior over a whole unlabelled test batch by iterating, for two rounds, three established operators: (i) k-nearest-neighbour (k-NN) label propagation on CLIP image features, (ii) lightweight LLM-free agriculture-aware prompts, and (iii) a Sinkhorn–Knopp doubly stochastic projection. On three agricultural fine-grained benchmarks (PlantDoc, VL-PAW, PP2021) with a single frozen OpenAI-CLIP ViT-B/16 backbone, ITR improves top-1 accuracy by +11.87, +10.64, and +9.31 pp (mean +10.60), outperforming recent training-free test-time methods—TDA, CuPL, WaffleCLIP, CALIP, the transductive EM-Dirichlet, and the multi-model OTFusion (CLIP + DINOv2). A central part of ITR is a graph-reliability gate that, measuring the image graph’s edge label-homogeneity per batch, applies the structural refinement only where the graph is reliable and otherwise leaves the prediction untouched; it contributes +1.66 pp mean (+4.87 pp on VL-PAW) at no cost to any dataset. At one iteration, the Sinkhorn step reduces to a transductive correction of CLIP’s class-marginal bias, identifying the source of the gain. These results show that a frozen-CLIP model can be turned into a competitive agricultural classifier at test time, with no training or labelled data, offering a practical route to open-vocabulary plant-disease and weed recognition under label scarcity.

Fruit detection for small datasets via adjustable anchor boxes and transfer learning

Fruit detection is a crucial task in plant phenotyping but remains challenging due to limited training data, high variability in fruit appearances across different growth stages, and occlusions that hinder accurate detection. To address these issues, we propose an Adjustable Anchor Box Detection Network with Transfer Learning (ADNet_TL) for robust fruit detection under unstructured conditions. Our approach leverages the backbone of an existing detector to extract discriminative fruit regions, integrates two adjustable anchor box mechanisms that align with dataset-specific characteristics, and employs transfer learning to boost performance on small target datasets effectively. A comprehensive analysis is conducted to assess the impact of training sample sizes in both source and target domains. Experimental evaluations on Strawberry, Tomato, and Multi-fruit datasets reveal that ADNet_TL outperforms both the standard ADNet and the classical Single Shot MultiBox Detector (SSD), with up to a 14% improvement in mean Average Precision (mAP). These results underscore the potential of ADNet_TL for practical applications such as fruit forecasting and selective harvesting in real agricultural scenarios.

An intelligent computer vision-based deep learning framework for automated detection, localization, and classification of fruit diseases in precision agriculture systems

Cbam-augmented ResNet for high-accuracy grape leaf disease detection

Deep Learning for In-Field Occlusion Handling and Real-Time Fruit Detection Under Dense Canopy Conditions

Robotic fruit harvesting in dense canopies remains challenging due to occlusion, variable illumination, and fruit-foliage similarity. This review synthesises recent deep learning-based detection systems, with particular focus on occlusion mitigation through multi-stage perception pipelines. The literature reveals that attention mechanisms and multi-scale feature fusion have emerged as dominant strategies for detecting partially visible fruits under leaf cover and overlapping branches. Surveyed studies report mean Average Precision scores ranging from 85–95% on orchard imagery, with YOLOv8 and Faster R-CNN serving as common benchmarks. Real-time feasibility on embedded hardware has been demonstrated across multiple systems, though the review identifies persistent gaps, including inconsistent reporting across lighting conditions and fruit maturity stages, as well as ongoing challenges in reducing false positives from visual clutter. By categorising trade-offs between accuracy, efficiency, and robustness, this review consolidates current knowledge and highlights directions toward reliable autonomous harvesting in complex agricultural environments.