Skip to content

YichenQiu/deepdesign.space

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

29 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

DEEP DESIGN

An image classifier for interior design styles. User can upload a photo of an interior space and see how much this image belongs to a certian interior design style.

Try it out at www.deepdesign.space

Styles currently being recognized: Bohemian, Coastal, Industrial, Scandinavian

Data Preparation

scrape_image.py: Script to scrape interior design photos from Google Images and save them to AWS S3 bucket. Images were labeled using search terms, e.g. all images scraped using search term "Bohemian interior design" were labeld as Bohemian.

Image_data/dedup_image.py: Deletes duplicate images from folder.

Image_data: Contains all 1800 images used to train the model.

Feature Extraction

All images were passed through Google's pretrained Inception V3 neural network. 2048 features were extracted for each image. Explore_models.ipynb shows the process of using 5-Fold Cross-Validation to choose the best performing model.

MODEL LOG_LOSS SCORE
LOGISTIC REGRESSION 0.492
RANDOM FOREST CLASSIFIER 0.883
GRADIENT BOOSTING CLASSIFIER 0.59
ONE HIDDEN LAYER WITH SOFTMAX CLASSIFIER 0.4

The final model is a mini neural network with one hidden layer (1024 features) and softmax classifier.

Model/train_inception.py: Contains script to train the model using keras.

inception.h5: Saved Inception V3 model.

inV3_last_layer.h5: Saved neural network with one hidden layer and softmax classifier.

Model Evaluation

Model performance by design style. The model seems to recognize some Coastal and Industrial images as Scandinavian. However, this may not be considered as missclassification because sometimes one room may contain a mix of multiple design styles.

Web App

run python app.py to start the web app on local host 5000.

About

An image classifier for interior design styles

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published