Abstract
Compared handcrafted descriptors and CNN-derived features for classifying 15 cake categories (1,800 images). Low-level statistics (color histogram, edge direction, co-occurrence) fed an MLP but plateaued at 31% accuracy, while PVMLNet feature maps (layer −5) coupled with an MLP achieved 90% test accuracy. Transfer learning by fine-tuning PVMLNet reached 80%, highlighting the importance of deep representations.
About
The dataset of 15 cake types (chocolate, tiramisu, cheesecake, etc.) is split 100/20 per class for train/test. Handcrafted descriptors—color histograms, edge direction histograms, grey-level co-occurrence matrices—are concatenated and normalised (mean-var, min-max, max-abs) before feeding an MLP. Despite tuning, performance stagnates around 31% due to intra-class variability. Switch to PVMLNet: intermediate activations from layers −1 to −7 are compared, with flattened layer −5 delivering 90% accuracy and converging in <100 epochs. Transfer learning replaces PVMLNet’s final layer with the trained MLP head, but full fine-tuning settles at 80%, still below the feature-extraction approach. Error analysis via confusion matrices flags persistent confusions (e.g., chocolate-mousse vs ice-cream cake) and guides future data augmentation ideas.
Key results
- 90%
- Accuracy
- 90%
- Neural Features
- 31%
- Low-Level Features
- 80%
- Transfer Learning
Key findings
- Color-histogram-only MLP stabilised near 21% test accuracy after 5,000 epochs; adding edge-direction histograms raised it to 31%.
- PVMLNet layer −5 activations flattened into an MLP delivered 90% test accuracy with <100 training epochs.
- Fine-tuning PVMLNet via transfer learning achieved 80% accuracy, 10 percentage points below the feature-extraction pipeline.
- Confusion analysis showed chocolate-mousse, apple-pie and tiramisù frequently misclassified as ice-cream, carrot or chocolate cakes because of visual similarity.
Challenges
- Limited training samples (100 per class) hinder full fine-tuning of PVMLNet without overfitting.
- Visual similarity between categories exposes the need for texture- and shape-aware augmentations.
- Handcrafted descriptors require tedious scaling/normalisation yet still underperform deep features.
Methodology
- 01Extract color histograms, edge-direction histograms and GLCM features for MLP baselines
- 02Normalise handcrafted features with mean-variance, min-max and max-abs scaling
- 03Compare PVMLNet intermediate activations (layers −1…−7) and flatten layer −5
- 04Train MLPs on handcrafted and CNN-derived features in runs ≤100 epochs
- 05Replace PVMLNet’s final layer with the trained MLP head for transfer-learning experiments
