Some details
The main goal of the project was to build a solution to identify a brand-specific style using images as a data source. Style is defined as a well-known color scheme, image composition and anything that may help one to affiliate an image with some brand. For example, with theBMW trademark. In other words, we were faced with the task of classifying images.
Challenges
- Сars in the images might
Solutions
- We used the Darknet YOLOv3 algorithm to detect cars in the images, removed the rectangle bounding machine/machines– replaced it with a fully transparent color,cut a randomrectangle if there were no cars in an image. It was necessaryto provide the dataset homogeneity. Otherwise, images with and without rectangles would be split into 2 different classesand it will spoil the classification.
- We have performed data augmentation and created 10 newimages from each input image with applying small transformations – rotating, scaling, horizontally flipping,filtering (changing brightness, contrast, colors of images).Now the dataset was rich enough to use it for training.
- We have built a model that steadily copes with its purpose. The accuracy value for ourclassification is 95% on the test data. In the future, with the appearance of new data (new BMWadvertising campaigns, for example), it will also be possible to train a model on these images and improve neural network accuracy.
Business value
Such a solution could help young specialists in the Product design area with making sure their ownprojects are not plagiarizing any unique traits of well-known brands and are distinctive enough on its own.
