Table 3 Data preparation and splitting.
Process stage | Description |
|---|---|
Augmentation Techniques | RandomRotate90, HorizontalFlip, VerticalFlip, BrightnessContrast, GaussNoise, MotionBlur, HueSaturationValue, RandomGamma, ElasticTransform |
Preprocessing Pipeline | Resizing, RGB conversion, denoising, contrast & color normalization, histogram scaling |
Image Resize | 224 × 224 pixels (resized from original size) |
Color Space Conversion | Converted from BGR to RGB using OpenCV for color fidelity |
Brightness Normalization | Standardized to a consistent mean-brightness histogram per image |
Noise Handling | Gaussian and motion noise reduced; synthetic noise applied for generalization |
Train-Test Split | Split into 70% train (8400 images) and 30% test (3600 images), stratified across all 6 classes |
Training Sample Size | 8400 images (1400 per class × 6 classes) |
Testing Sample Size | 3600 images (600 per class × 6 classes) |
Train Class Balance | Even class distribution (Malabar, Red, Water × Fresh/Non-Fresh) |
Test Class Balance | Preserved balance across six classes in unseen data |
Train RGB Statistics | R: 180–210, G: 175–205, B: 160–200 (mean ± 10); normalized between 0–1 |
Test RGB Statistics | R: 175–205, G: 170–200, B: 160–195 (mean ± 10); normalized between 0–1 |
