Lab notebook · 08 / 27Model comparison

Model-selection study on flower images

How do a small CNN and two pretrained feature extractors compare on the same flower-image task?

University coursework

Scope of this result

One validation split selects and scores the winner. This compares complete approaches; it is not an independent test result.

The evidence

Fig. 01

Bar chart of best validation accuracy: EfficientNetB0 88.6%, MobileNetV2 87.9%, and scratch CNN 60.2%; alongside eight-epoch validation learning curves. — One disjoint 80/20 split and eight epochs for each complete approach. Frozen EfficientNetB0 reaches 88.6% validation accuracy, frozen MobileNetV2 87.9%, and the scratch CNN 60.2%.

01 · Method and interpretation

Three image classifiers share one deterministic 80/20 split, 160px inputs, augmentation, and an eight-epoch budget on 3,670 flower photos: a small CNN trained from scratch and frozen ImageNet feature extractors using MobileNetV2 and EfficientNetB0. EfficientNetB0 reaches 88.6% validation accuracy, MobileNetV2 87.9%, and the scratch CNN 60.2%. Because architecture, pretraining, and trainable capacity change together—and the same validation split selects and scores the winner—this is a practical model-selection benchmark, not a controlled architecture ablation or unbiased test estimate.

02 · Supporting figures

Selected outputs from the complete notebook.

Fig. 02

Confusion matrix for EfficientNetB0 on 734 validation images across daisy, dandelion, roses, sunflowers, and tulips. The diagonal sums to 650 correct predictions, or 88.6%. — The best-epoch EfficientNetB0 confusion matrix contains 734 validation images. Roses and tulips account for the largest off-diagonal pair.

Fig. 03

Eight sample training photos labeled roses, tulips, sunflowers, and daisy, at natural photo resolution with real backgrounds. — The task: five flower species at photo resolution, messy backgrounds and all.