Automatic Fruit Quality Classification System: A Comparative Study Between Traditional Machine Learning and Deep Learning

This article introduces a computer vision-based automatic fruit quality classification system, aiming to automate fruit quality detection and grading in agricultural industrial scenarios. The system can identify the commercial quality grades (good/average/poor) and size categories (small/medium/large) of fruits. By comparing the performance of traditional machine learning models (SVM, XGBoost) and deep learning models (CNN), it finally proposes a hybrid architecture scheme suitable for industrial scenarios. The original author team of the project is karoldmejia et al., the source is GitHub, and the release date is 2026-06-06.

Project Core Objectives: 1. Classify fruits into three commercial grades (good, average, poor) based on appearance features; 2. Divide into small/medium/large size categories based on pixel area; 3. Compare the performance of traditional ML and deep learning models; 4. Analyze the impact of geometric and color features on the models; 5. Propose an industrially feasible scheme. Dataset Situation: Contains images of six types of fruits: apples, bananas, guavas, lemons, oranges, and pomegranates, with a total of 36848 samples. Quality labels are manually annotated, and size labels are divided based on the normalized pixel area of fruits using type-specific percentiles (small <33%, medium 33-66%, large>66%) to ensure classification consistency.

Image Processing Stage: 1. YOLOv8 object detection to locate fruit regions; 2. HSV color space analysis to enhance contrast; 3. Contour detection to extract boundaries; 4. Watershed algorithm to handle overlapping fruits; 5. Filter invalid samples and uniformly adjust images to 224×224 pixels; 6. Apply data augmentation (rotation, flipping, brightness adjustment, etc.) to minority class samples. Feature Extraction: Geometric features (pixel area, aspect ratio, coverage rate) are used for size classification; color features (first-order statistics in HSV space such as mean/standard deviation, second-order statistics such as texture features and color consistency) are used for quality classification.

Traditional Machine Learning Models:

• SVM: A classic kernel method that uses geometric and color features to construct an optimal hyperplane for classification.
• XGBoost: A gradient-boosted decision tree ensemble method that automatically captures non-linear interactions of features, has built-in regularization to prevent overfitting, and is suitable for tabular features. Deep Learning Models:
• CNN: Directly learns hierarchical visual features from raw images through convolutional layers, pooling layers, and fully connected layers in an end-to-end manner, without the need for manually designed feature extractors.

Quality Classification Results (macro-average F1-Score): SVM 0.9319, XGBoost 0.9494, CNN 0.9497 (best). Analysis: CNN excels at recognizing complex surface patterns (defects, textures), making it suitable for quality assessment. Size Classification Results: SVM 0.9590, XGBoost 0.9813 (best), CNN 0.9181. Analysis: XGBoost effectively utilizes explicit geometric features, making it suitable for size classification.

Core Findings: Different models have advantages in specific tasks—CNN is suitable for spatial information/texture-related tasks (quality), while traditional models are suitable for tasks involving explicit quantitative features (size). Hybrid Architecture: Input image → YOLOv8 detection → Feature extraction → Branch to XGBoost (size classification) and CNN (quality classification) → Comprehensive output. Advantages: Maximized accuracy (XGBoost size F1=0.9813, CNN quality F1=0.9497), optimized efficiency, interpretability (XGBoost feature importance), and independent module updates.

Practical Applications: Automation of fruit sorting lines, quality grade determination, packaging specification allocation, yield statistics, and quality control. Economic Benefits: Reduce labor costs, improve consistency, accelerate processing speed, and reduce losses. Technology Stack: OpenCV (image processing), NumPy (numerical computation), Pandas (data processing), Scikit-Learn (SVM), XGBoost (gradient boosting), TensorFlow/Keras (CNN), Albumentations (data augmentation), YOLOv8 (object detection), Matplotlib (visualization).