idk100/intel-image-classification-cnn-comparison

Intel Image Classification — Trained Checkpoints

Eight trained Keras 3 checkpoints from a systematic comparison of CNN architectural patterns on the Intel Image Classification dataset (6 scene classes: buildings, forest, glacier, mountain, sea, street).

Full methodology, code, and discussion: GitHub repo

Also on the Hub: idk100/intel-image-classification-cnn-comparison

Checkpoints

File	Description	Test Accuracy	Test Loss
Finetuned_Xception41.keras	Pretrained Xception-41 backbone, fine-tuned	0.910	0.367
Pretrained_FrozenHead.keras	Pretrained Xception-41 backbone, frozen, new head only	0.892	0.359
XCeption_Lite_V2.keras	From-scratch Xception-style (8x middle flow) + dropout + augmentation	0.870	0.388
Base_CNN_Model.keras	Baseline CNN, no architectural tricks	0.856	0.425
Batch_Normalization_Added.keras	Baseline + batch normalization	0.854	0.430
Depthwise_Separable_Added.keras	+ depthwise-separable convolutions	0.840	0.466
XCeption_Lite_V1.keras	From-scratch Xception-style (4x middle flow), no regularization	0.828	0.451
Residual_Added.keras	Baseline + residual connections	0.827	0.509

Loading a checkpoint

from huggingface_hub import hf_hub_download
import keras

path = hf_hub_download(
    repo_id="idk100/intel-image-classification-cnn-comparison",
    filename="Finetuned_Xception41.keras",
)
model = keras.models.load_model(path)

Input preprocessing differs by model: the five from-scratch checkpoints (Base_CNN through XCeption_Lite_V1/V2) expect 150x150 RGB input rescaled to [0,1]. The two pretrained-backbone checkpoints (Pretrained_FrozenHead, Finetuned_Xception41) expect 299x299 RGB input normalized via keras_hub's Xception-41 ImageConverter preset, not simple [0,1] rescaling.

License

MIT — see the GitHub repo for details.