Fruit Quality Image Classification

• Fruit quality image classification is the automated evaluation of produce using digital imaging and machine learning to replace manual inspection.
• It integrates classical feature extraction methods and modern deep learning architectures to accurately estimate ripeness and detect defects using metrics like accuracy and IoU.
• Practical deployments on mobile and embedded devices enable real-time grading and scalable assessment in agricultural supply chains.

Fruit quality image classification refers to the automated assessment, grading, and defect detection of fruit samples using digital images and advanced machine learning models. The goal is to replace or augment labor-intensive manual sorting and visual inspection in both postharvest and supply-chain contexts with scalable, objective, and reproducible computer vision solutions. The domain encompasses ripeness estimation, defect identification (e.g., disease spots, bruising, deformities), and fine-grained market grading. Performance is typically quantified via metrics such as accuracy, precision, recall, F₁-score, and intersection over union (IoU), with state-of-the-art systems leveraging architectures from classical machine learning with engineered features up to deep learning with transfer learning and vision transformers.

1. Datasets, Labeling Protocols, and Preprocessing

Image datasets for fruit quality classification range in size from a few hundred (legacy feature-based or disease datasets) to tens of thousands of images, curated across diverse fruit types and agricultural conditions. Representative examples include CASC IFW apples (5,858 images, binary healthy/worm spot) and Banana Fayoum ripeness (273 images across four ordinal classes) (Knott et al., 2022), FruitNet (19,526 images, six fruits × three quality grades) (Morshed et al., 2022), and DragonFruitQualityNet’s dragon fruit corpus (13,789 samples across fresh, immature, mature, defective) (Haquea et al., 10 Aug 2025).

Labeling protocols emphasize multi-class and fine-grained annotation. For instance, lychee datasets use a three-class maturity label: unripe, semi-ripe, ripe (Zhang et al., 19 Oct 2025). Disease and defect datasets typically require bounding-box or mask annotation for each affected region, validated by multiple reviewers (Zhang et al., 19 Oct 2025). Preprocessing includes resizing (256×256 or 224×224 for CNN input), normalization (e.g., scaling RGB to [0,1] or mean/std per channel), and on-the-fly runtime data augmentation (random flips, rotations, contrast/brightness jitter) to counter class imbalance and improve model generalization (Morshed et al., 2022, Haquea et al., 10 Aug 2025).

Synthetic image augmentation using generative AI (MidJourney, Firefly, cGANs) further expands dataset diversity, with realism quantified by PSNR and SSIM to validate structural fidelity (Yoon et al., 15 Jul 2024, Bird et al., 2021).

2. Classical Feature-Based and Shallow Learning Approaches

Legacy fruit quality systems rely on manually engineered features for color, texture, and shape:

• Color histograms: Computed in RGB, HSV, or CIELAB space over segmented regions; for example, GCH and CCV (Dubey et al., 2014, Dubey et al., 2014).
• Texture descriptors: Local Binary Patterns (LBP), Local Ternary Patterns (LTP), Completed LBP (CLBP), structure-element histograms, and GLCM (Dubey et al., 2014, Dubey et al., 2014, Rizzo et al., 2022).
• Shape metrics: Area, roundness, aspect ratio, Hu moments (Rizzo et al., 2022).

Defect localization usually precedes feature extraction via k-means clustering in perceptual (Lab or HSV) space, optimizing cluster number and thresholds to isolate candidate regions (Dubey et al., 2014, Dubey et al., 2014). Feature vectors are concatenated and normalized, then supplied to multi-class SVMs (Gaussian RBF kernel, one-vs-one coding) or other shallow classifiers (k-NN, random forests).

Performance with these approaches typically peaks around 90–93% accuracy for apple disease/defect classification, with engineered feature fusion yielding up to +10 percentage points over color or texture cues alone (Dubey et al., 2014, Dubey et al., 2014).

3. Deep Learning Architectures and Training Paradigms

Modern fruit quality image classification increasingly favors deep convolutional architectures pre-trained on large general corpora (ImageNet):

• CNN backbones: DenseNet201, ResNet-18/50/152, VGG16, MobileNetV2, Xception, EfficientNetV2-B0, InceptionV3 (Morshed et al., 2022, Darapaneni et al., 2022, Han et al., 27 Feb 2025, Peón et al., 31 Jul 2025).
• Vision Transformers: Self-supervised DINO ViTs (ViT-S/8, ViT-B/8) as frozen feature extractors paired with shallow classifiers (SVM, XGBoost, MLP) (Knott et al., 2022).
• Multi-input architectures: RGB and silhouette image branches (segmented via Segment Anything Model or Otsu thresholding), fused with MLP heads for defect and deformity detection (Chuquimarca et al., 14 Oct 2024, Beltran et al., 17 Dec 2024).

Transfer learning is ubiquitous, with early layers frozen to preserve domain-invariant low-level knowledge and top layers fine-tuned. Typical input dimensions are 224×224 or 256×256, and loss functions are categorical cross-entropy (multi-class) or binary cross-entropy (defect presence). Data augmentation, dropout, and batch normalization mitigate overfitting.

Cutting-edge results demonstrate DenseNet201 + augmentation achieving 99.67% accuracy (six fruits × three grades) (Morshed et al., 2022), MobileNetV2 multi-input reaching 100% (apples) and >92% (mangoes, strawberries) for deformities (Beltran et al., 17 Dec 2024, Chuquimarca et al., 14 Oct 2024), and ResNet-18/50/101/152 attaining >90% for ripeness and disease classification in mangoes (Peón et al., 31 Jul 2025). Shallow CNNs and classic transfer heads yield lower but still robust accuracy (>85%) for tasks such as palm fruit maturity (Han et al., 27 Feb 2025).

4. Generative Augmentation and Data Scarcity Solutions

Synthetic image generation with text-to-image or image-to-image diffusion models (Easy Diffusion, MidJourney, Firefly), as well as class-conditional GANs, is deployed to overcome dataset sparsity and class imbalance (Yoon et al., 15 Jul 2024, Bird et al., 2021, Beltran et al., 17 Dec 2024). Metrics such as PSNR (27–29 dB) and SSIM (up to 0.42) validate synthetic image realism, particularly in post-harvest scenarios with critical surface details.

Augmenting training data with synthetic images demonstrably boosts classifier performance. In lemon quality, cGAN augmentation raises VGG16 accuracy from 83.77% to 88.75% (+4.98 pp) (Bird et al., 2021). Mixing real and synthetic melon images increases YOLOv9 recall/mAP by ~3–5% (Yoon et al., 15 Jul 2024). Optimal ratios (~1:1) help avoid domain bias.

Grad-CAM and other explainability methods demonstrate that synthetic images retain class-discriminative cues utilized by CNN classifiers for improved defect detection (Bird et al., 2021).

5. Specialized Detection, Grading, and Quality Metrics

Object detection pipelines based on YOLO (v3/v8/v9) and RT-DETR architectures quantify not only presence but precise localization, enabling per-fruit grading for traits such as ripeness, disease, and net quality (Nagpal et al., 2023, Yoon et al., 15 Jul 2024, Zhang et al., 19 Oct 2025). Detection heads predict bounding boxes and class probabilities in a single forward pass; grading is accomplished either as a secondary head or through feature regression.

YOLO-based systems achieve close to 99% detection accuracy and >90% IoU for cherry counting, size, and color estimation, with speed and consistency exceeding manual evaluation (Nagpal et al., 2023). For lychee and melon, [email protected] ≥ 0.98, precision/recall ~0.98 are typical with data augmentation (Zhang et al., 19 Oct 2025, Yoon et al., 15 Jul 2024).

Specialized quality metrics—such as net density ($\rho_{\text{net}}$) and uniformity ($\sigma(A_i)$) for melons, or deformity class for segmented apples/mangoes/strawberries—enable grade assignment according to market standards. Multimodal fusion strategies (RGB+depth, RGB+HSI) are anticipated to further boost discrimination in challenging scenarios (Zhang et al., 19 Oct 2025, Rizzo et al., 2022).

6. Practical Deployment, Industrial Integration, and Limitations

Deployment recommendations emphasize lightweight, low-resource inference for small-scale, decentralized stakeholders:

• ViT feature extraction for smartphone-based image pipelines; shallow classifiers run on CPU-only hardware—90% accuracy with 3× fewer labeled samples than CNNs (Knott et al., 2022).
• Real-time grading on mobile devices via optimized CNNs (DragonFruitQualityNet <31M params, TFLite quantization to 7MB, 30–50ms/img inference) (Haquea et al., 10 Aug 2025).
• Graphical interfaces (MATLAB App Designer) streamline multi-stage detection-classification workflows for farm automation (Peón et al., 31 Jul 2025).

Sample efficiency and robustness to drift are key, with unsupervised visualizations (PCA, UMAP) confirming better class separability in deep and transformer embeddings for subtle defects (Knott et al., 2022). Limiting factors include reliance on upstream segmentation quality (SAM failure cases), data bias toward controlled environments, sensitivity to out-of-distribution lighting and occlusion, and the absence of localization in pure classifiers (Chuquimarca et al., 14 Oct 2024, Knott et al., 2022).

7. Research Trends, Benchmarking, and Future Directions

Emerging research themes include:

• Vision transformer architectures (ViT, DINO-ViT) for improved sample efficiency and transparency; self-supervised feature learning on unlabeled corpora (Knott et al., 2022, Rizzo et al., 2022).
• Explainable AI (Grad-CAM, LIME, SHAP) for regulatory contexts and operational trust (Bird et al., 2021, Rizzo et al., 2022).
• Synthetic augmentation via advanced GANs and diffusion models controlling multi-attribute severity (Yoon et al., 15 Jul 2024, Bird et al., 2021).
• Multimodal and multi-task learning for simultaneous trait grading (ripeness, firmness, disease) (Rizzo et al., 2022, Zhang et al., 19 Oct 2025).
• Edge deployment on embedded GPUs and smartphones, latency benchmarking, and quantized models for sub-100KB footprint (Haquea et al., 10 Aug 2025, Rizzo et al., 2022).

Open benchmarking, diverse multi-variety datasets with robust annotation and detailed splitting protocols are recognized as essential for progress and cross-paper comparability (Morshed et al., 2022, Zhang et al., 19 Oct 2025, Rizzo et al., 2022). Researchers recommend progressive fusion of spectral, depth, and traditional RGB channels for enhanced discriminatory power, especially on marginal or occluded samples.

The field is moving toward universal, sample-efficient, explainable, and highly automated fruit quality image classification pipelines, spanning detection, grading, and trait prediction suitable for both industrial and resource-constrained agricultural contexts.