carl-cnn's Introduction

CarL-CNN

Car Logos CNN - building my own car logos classifying neural network

Description

CarL-CNN was trained with 20,778 50x50px RGB images depicting logotypes of 40 different car brands. The dataset was handpicked from loosely scraped websites and contains images, photos, drawings, sketches at various color schemes (black/white, RGB, CMYK, one-color), different angles and can sometimes contain some noise (other logotypes, background, etc.)

It classifies a given image by attributing a predicted brand label:

Model Metrics

The model got the below scores:

Precision: 94.20%
Recall: 94.03%
F1 score: 94.04%
Accuracy: 94.03%

Features

The Jupyter Notebook file contains method definitions for showcasing, class probability prediction and new image recognition. Also, a closer examination of the wrongly predicted cases is done to analyze which car brands need some dataset enrichment.

Future development

I plan to make it into a web application and enable adding user-owned images for classification. Later on, using sl4a, I plan to make it into an Android application and enable car logo recognition by mobile camera -- a True computer vision!

Links - needed for the full run of CarL-CNN

(unzip to folder, where CarLogosCNN.ipynb resides)

Model weights (1 file, 140 MB) - http://u.42.pl/GEt0_model_weights
Images (20778 files, 54 MB - ZIP) for the train and test set - http://u.42.pl/GEt7_images
New images (12 files, 677 kB) unseen during the training - http://u.42.pl/GEtf_new_images

Requirements

numpy
matplotlib
scikit-learn
keras (tensorflow backend)
pillow - for image processing

carl-cnn's People

Contributors

Stargazers

Watchers

carl-cnn's Issues

Images for CarL-CNN

Question about Test Results

hi, kuba, thank u for your code, in the past few days, I have trained with your code, the data set is the same, but the best precision of the model is only 84%.epoch=200 and train_size=0.8,It's very different from your results, so I want to know if you have improved the model or added other data sets.thank u again!

Reason of ImageDataGenerator

Hi @kuba-siekierzynski ,

i would like to thank you for sharing this code , looks promising and good results till now .
i just want to discuss with you the reason of ImageDataGenerator . i understand that it increase accuracy but how and have you tried any alternatives , also you fit the training data only and don't do the same on the test data .

# In order to process data augmentation, keras' ImageDataGenerator can apply some transformations to images on random
from keras.preprocessing.image import ImageDataGenerator

datagen = ImageDataGenerator(
    # featurewise_center=False,
    # samplewise_center=False,
    # featurewise_std_normalization=True,
    # samplewise_std_normalization=True,
    # zca_whitening=True,
    rotation_range=45,
    width_shift_range=0.2,
    height_shift_range=0.2,
    horizontal_flip=True,
    vertical_flip=True)

i will keep playing on this model and try to raise the accuracy on some images that fails as steering wheels ones

kuba-siekierzynski / carl-cnn Goto Github PK