Our solution

Software module that can analyze the large libraries of images and correctly tag them based on their content. In the given case it was not possible to use out-of-the-box solutions due to custom tags and a requirement of having a very limited amount of tags per image.

Results

Our model achieved an accuracy of ~95%.

Benefits for the customer

The customer was able to further enhance their web services and provide better filtering options for the users.

Details

• We built here a Convolutional Neural Network (CNN / ConvNet) – a special type of feed-forward networks. CNNs are known as a very well method for solving visual recognition tasks.
• To train our model, we employed the transfer learning technique that utilizes a trained model on a different dataset, and adapts it to our problem.
• To implement the network we used the Caffe framework. The training process was executed using cluster equipped with NVIDIA Tesla K80 GPUs.