Computer Vision News - November 2016

The other problem is having packages of different sizes that exist for the same product. If you take a photo of one specific brand, many times the labels of the different sizes are very similar. For example, you might have one product that is 200 mg and another that is 2 kg. When the computer sees the packages, it might think they look exactly the same. On the contrary, it must be able to separate them to recognize these differences. With Yodigram, it doesn’t just recognize the labels, it also detects the exact item which must be identified. Packaging is also an issue, when it is not rigid: it can come in many different shapes and positions, even when it is the same product. Vasileiou and his team have tried many approaches to these challenges. Now they are using a hybrid deep learning approach because they do not consider deep learning good enough for detection: you must give segments to the deep learning system to be able to separate the products. On the other hand, they use deep learning for classification, and have designed their own model for this purpose. When you are very close, Vasileiou explains, you have many less distortions, but in each environment it is different. That means you must be able to extract features that are different in each environment or train the system to be invariant to these problems. He was able to train the systems using data from the pilot project conducted with a local customer who has run the system for almost a year; that enabled him to develop a lot of data and make the system robust enough. The pilot is expected to be completed by end of September, just before our readers read this report. After that, with full confidence that the system works in many different environments and with many different packages, Yodigram plans to expand. They have started working with the UI, to be able to work in multiple languages. The system does not give much weight to OCR since it doesn’t rely on the meaning of what the label says. It uses the features of the labels “more in a deep learning sense”. Computer Vision News Application 67 “Which items are exposed on the shelf in store” Application Product Recognition Interface