Blog DataArt Research Lab to experiment on finding and proofing feature extraction methods suitable for food recognition tasks

Event Image
Meals named the same rarely look similar. This is not only due to different people cook differently – in the computer vision sense, meals are combinations of areas (spots) with different color, texture, shape each. This makes typical image recognition principles less suitable for food image recognition, as we cannot rely on either form or relative position of the image parts. Typically, if local peculiarities of objects being detected cannot be caught, integration feature extraction methods take over differential one – e.g. in our current food image classification engine we mostly rely on combined histogram and texture parameters for the whole image. This approach shows relatively good results unless the meal we’re trying to classify appears to have no noticeable texture features.

In this case, the classification decision is based on the histogram output. If we then remember tough that the histogram is just the measure of quantity of one color over another, we understand that the histogram output would be the same for, say, a mix of diced or whole vegetables – as both have the same color proportion over the whole image.

To overcome this while preserving invariance to scale, rotational and spatial differences, a new method, called ‘color spots distance’ or just ‘color distance’ was invented and explored. The method is based on the color histogram with the extension that calculates a measure of proximity of the spots of a certain color over the image. This measure accounts not only for the color itself, but also for the degree of color distribution, thus giving different results for, say, a whole tomato (where all red components are compact) against a sliced one, where they are distributed.

The new method has been proved against the sample set of 15 fruits of different kind, showing the recognition quality of about 90% ‘in vitro’. Field experiments following the lab tests are to follow, yet the preliminary results are quite encouraging. To compare – the texture method referred above shows only as little as 40% under the same conditions.

The test results chart and misclassification graph for the color distance method are given in the next post.

Subscribe to our news

One thought on “DataArt Research Lab to experiment on finding and proofing feature extraction methods suitable for food recognition tasks”

Leave a Reply