I am using Deep neural networks to classify food images by cuisine. I am changing the parameters, inputs to get maximum accuracy. I am finding that just using the blue component of the images (a greyscale of blue - I am not changing R and G to 0) gives me better accuracy than RGB. I am trying to find out why this is happening.
Any ideas or relevant literature is greatly appreciated!
From your description, it is not perfectly clear what you are doing: I have to guess a bit that you're using the "blue excerpt" from the RGB image to classify the images.
As why blue might outperform, my guess is more educated:
- green and red prevail in food. For the same item you may get a lot of green and/ro red shades - not perfect for classification.
- blue is comparably rare. And the shade of blue of some item does not vary that much depending on the state of ripeness etc.
So if some shade blue is available at all, it should make up for a more constant criterion for classification. Thus...
From your description, it is not perfectly clear what you are doing: I have to guess a bit that you're using the "blue excerpt" from the RGB image to classify the images.
As why blue might outperform, my guess is more educated:
- green and red prevail in food. For the same item you may get a lot of green and/ro red shades - not perfect for classification.
- blue is comparably rare. And the shade of blue of some item does not vary that much depending on the state of ripeness etc.
So if some shade blue is available at all, it should make up for a more constant criterion for classification. Thus...
Ah, thank you! I am using the blue excerpt, so my input is image size x 1.
I printed the image out, and saw that blue excerpts has a higher 'contrast', for lack of a better term.
I assumed that red and green would perform better, since some cuisines have more 'redness' in their food. Your guess makes more sense. I will need to dig deeper into this. Thanks again!
Hi, I understood what you wanted to say. I have experimented with this color components. I have published a paper on this (only red channel) and another paper is in review with details analysis on various color channels separately (R /G/B)along with RGB.
I have experimented on medical images.
But I also found out blue channel sometimes perform better than RGB, but sometimes not. Its not always true that blue channel performs better than RGB. But yes one thing is true, both RGB and blue channel outperformed red and green channel in terms of accuracy and AUC.
Why it happens you can have a research on them or ask some image processing expert who might have better answer.
Indeed mr. Dreher is correct. In general when using a classification criterion for A- non A categories the most useful one is the simplest association that is common for one category and ideally absent or rare in the other . As he observed, blue is rare in food that means the "transfer function" that the network must internally reach is.. blue= non-food (+some noise) , no blue= food. If the network is trained with altered image sets showing mostly food under blue lighting or blue food ( if there is such a thing) mixed with other pictures that have some blue in them , the network is forced to guess more convoluted associations and the initial results will be catastrophic until it finds that (mostly red + yellow and low blue or blue overall and red and yellow= probably food).
The training set is crucial to neural network performance. The story about poorly trained neural networks that made the military reject a tank -detecting automated image analysis software is well known.. it goes that some 25 years ago they tried to analyze images from UAVs to automatically determine if tanks were present or not- using a neural network (its exact architecture is not known to me but I think it was perceptron-type). The network received training using pictures that had tanks in them, some well camouflaged some plain in the open field, etc and showed a good in-the-lab response. On the field it was the opposite- bad performance throughout, so much that even when in desperation the UAV was flown over an entire field packed with non-camouflaged ,surplus tanks none was detected. Later it was determined that the training image set was faulty: all the tanks in the pictures were photographed on sunny days, and the pictures with no tanks in them ..were shot on overcast days. As the weather was rather cloudy in the days of the test....it finally dawned on our researchers that after some millions in expenditures on novel technology they achieved a novel, neural-network enabled flying sunny weather detector.
The fault was later fixed but the brass lost interest...
- Badges
- Science topic
- Similar topics
- Computer Science
- Images