README.md

metadata

title: Intel Scene CNN
emoji: 🏔️
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 5.9.1
app_file: app.py
pinned: false
license: mit

Intel Scene Classification: a CNN built from scratch

A plain convolutional neural network (no ResNet, no pretrained weights) trained on the Intel Image Classification dataset: ~25k natural scene photos in 6 classes: buildings, forest, glacier, mountain, sea, street.

This is the follow-up to my EuroSAT satellite CNN. It tests whether the same architecture generalizes to street-level photos, where orientation matters and the classes overlap visually.

Live demo: upload an image, get the predicted class and a Grad-CAM heatmap, plus the test metrics.

Results

Metric	Value
Test accuracy (seg_test, 3,000 images)	89.87%
Macro F1	0.900
Best validation accuracy	90.98%
Parameters	1,206,886

Class	Precision	Recall	F1	Test images
buildings	0.913	0.860	0.886	437
forest	0.979	0.985	0.982	474
glacier	0.846	0.866	0.856	553
mountain	0.876	0.830	0.852	525
sea	0.895	0.937	0.916	510
street	0.897	0.918	0.907	501

Where it fails: mountain → glacier (65), buildings → street (46), glacier → mountain (46), street → buildings (29). Almost all errors come from two pairs: snow-covered peaks vs. glaciers, and urban scenes where a photo of a street is also a photo of buildings. Forest is near-perfect (98% F1) because its texture and color are unique. Several high-confidence "mistakes" are labeling errors in the dataset itself.

Dataset

Split	Images
seg_train → train (85%)	~11,900
seg_train → validation (15%, stratified)	~2,100
seg_test (held-out test)	3,000
seg_pred (unlabeled)	~7,300

Architecture

flowchart LR
    A[Input 3×128×128] --> B1[Block 1<br/>32ch → 64×64]
    B1 --> B2[Block 2<br/>64ch → 32×32]
    B2 --> B3[Block 3<br/>128ch → 16×16]
    B3 --> B4[Block 4<br/>256ch → 8×8]
    B4 --> G[Global Avg Pool<br/>256]
    G --> F1[Dropout 0.4 → FC 128 → ReLU]
    F1 --> F2[FC 6 → softmax]

Each block is Conv3×3 → BN → ReLU → Conv3×3 → BN → ReLU → MaxPool 2×2 → Dropout2d(0.1).

Stage	Output shape
Input	3 × 128 × 128
Block 1	32 × 64 × 64
Block 2	64 × 32 × 32
Block 3	128 × 16 × 16
Block 4	256 × 8 × 8
Global average pool	256
FC + ReLU	128
FC (logits)	6

Training setup

• Optimizer: AdamW (lr 1e-3, weight decay 1e-2), cosine LR schedule, 30 epochs
• Loss: cross-entropy with label smoothing 0.1
• Image size: 128×128, normalized with per-channel stats computed from the training data
• Augmentation: random resized crop (0.75–1.0), horizontal flip, ±10° rotation, mild color jitter
• No vertical flips or large rotations. Scenes have gravity (sky on top), unlike satellite images.
• Model selection: best validation accuracy checkpoint

What the CNN learns

Feature maps before vs. after training

First-layer filters before vs. after training

Grad-CAM: where the network looks

Confidence and mistakes

Run locally

pip install -r requirements.txt gradio
python app.py

Files

File	Purpose
app.py	Gradio app: upload, prediction, Grad-CAM, metrics, plots
model.py	SceneCNN architecture + Grad-CAM
cnn_intel.pth	Trained weights (state_dict)
cnn_intel_meta.json	Class names, normalization stats, width, image size
metrics.json	Test metrics shown in the app
images/	All plots from training and evaluation

Author

Chandra Prakash Bathula, Adjunct Faculty, Saint Louis University (School for Professional Studies)