[WEEK 4–5] - Brand Recognition Using Machine Learning from Google Street View Images

Theme: Predicting the Brands

Team Members : Alper Fırat Kaya, Hülya Şermin Karakaş

We didn’t write for a while because of the homework density from all other courses but we have lots of exciting news! Finally we finished the Brand Recognition project and the results are much better than we predict from the beginning of the project.

On those two weeks, we worked on the localization, classification and recognition parts of the project. The most challenging an informative part of the project is the recognition part of the project. We used multilayer perceptron known as multilayer neural network on the project with using Char 74k dataset.

The dataset contains 74k images but we used the 25k of them which are related to the natural image based approach. The 75 percent of the images used for the training and the 25 percent of the images used for validation.

For the MLP, we used 2 hidden layer and we got %83.16 accuracy from from validation dataset.

For the test part, we collected the Google Street Images manually from the web worldwide. According to our experiment, we got %55.62 accuracy rate.

It was a great opportunity to work on such a project about cities and machine learning. We learned a lot about scene text recognition, localization/classification of natural images and machine learning libraries.