There are a lot of algorithms -especially in areas like image processing that *seem* like they work really well. Sometimes I wonder if this is because they have been tested on images where the results are really quite impressive. Take for example the task of haze suppression in photographs. I have run a couple of algorithms on some images with real haze in them and to be honest the results are quite good. The photographs were taken on Monte Generoso in the Swiss Canton of Ticino. Here’s the first result. Clearly the background haze in the valley has been suppressed, but not totally removed. The colours are more saturated and vibrant – but do they now lack natural colour?
This algorithm runs reasonably quickly on the original at 3000 x 4000 pixels – 252 seconds – that may seem like a long time, but these algorithms are intrinsically complex. Too complex for a mobile device. Let’s try algorithm number two. Here’s the result. Is it better? The colours certainly don’t seem as over-saturated, and it appears as though more haze has been removed.
The problem here is speed. I down-sampled the image here to 600×800, and the running time was 112 seconds. For a 900×1200, 364 seconds. You can see where this is going. The larger the image, the more it struggles to get the processing done. Both algorithms were written in Matlab. As a second example, consider this image, again from the same area – haze was particularly bad that day.
The first image shows the result of processing with Alg.1, the second with Alg.2. The first algorithm seems to have left some hazy artifacts near the barn and rocks on the hillside. The second algorithm seems to have suppressed the haze better, and although being a little darker retains more of the image detail. The caveat – 433 seconds on a 900 x 1200 image.
To make these into a competent app, or even a “filter” on a camera would require some incredulous increase in efficiency. Algorithm 2, which produced the better visual results would take over 100 minutes to process the original 3000 x 4000 pixel photograph.
Beauty of course is truly in the eye of the beholder.