Digital Camera Could Help Sort Fish, Save Stocks

MountainSplash writes "PlanetArk.com is carrying a story about a new camera that "takes a digital photograph of the catch which is then divided into a grid, allowing a computer to measure the shape and color of each fish in the grid. It needs one tenth of a second and identifies 98 percent of fish correctly." The claim is that fish can then be culled quicker possibly increasing the likelyhood of survival for the incidental catch in the net. Testing is being done by Norway's Institute of Marine Research and Norwegian marine electronics maker Scantrol. Onboard testing has proven highly successful, but underwater attempts still need more work. With everything we have all been seeing computers do the last few years, I personally found this to be one of the more interesting of late."

Comments

Pattern recognition

[erixtark]

This is just one example of the increased power of automated pattern recognition. Once computers reach a level of vision close to humans, we will se an explosion in automated tasks. Other examples include Sony Aibos vision, lip-reading software that helps in speech recognition, 'robot scientists' and the next generation of speech recognition with the potential to revolutionize human computer interaction. HAL, is that you?

Nice, but typical marketing optimism...

[ControlFreal]

IAACVPS (I Am A Computer Vision Ph.D. Student), and I'd like to add some general remarks concerning this application, and concerning computer vision in general.

Although the article mentions a nice application of computer vision, it is shockingly sparse in details. This in itself is not so strange for a news-site, but the fact that they didn't include a link to a more detailed description is a pity.

Some ideas:

First, the article doesn't make sufficiently clear whether one looks at the net, full of fish, or that one looks at the fish all spread out on a flat surface. If one looks at a full net, one can only see the fish on the outside, i.e., only a small fraction: that doesn't provide any information on the fish on the inside. If one looks at the fish spread out on a flat surface, one can see all the fish, but there are a number of issues here:

• Orientation-variance: if a fish is lying head-down (e.g. because it is crammed a bit too tightly between other fish), it's hard to measure its size.
• Occlusion: Even on a flat surface, on fish' tail might cover part of another fish' body. This makes measuring difficult.
• Lighting-variance: Fish are shiny reflective critter. So some parts of the image might have very white spots in them. This makes the application of automated algorithms difficult. Make no mistake: we as humans can recognize objects almost irrespective of the intensity and color of the light, but computer vision algorithms have severe difficulties doing this!
• Determination of number: How, exactly, do they count the fish, given all the difficulties listed above? So: what does that 90-something% accuracy mean?

Given the speed at which they process, it's most likely that they determine fish-size based on general statistical properties in different regions of the image. In that case, the 90-something% accuracy really doesn't mean that much, because in all honesty, I don't see how they can either measure or guarantee that. Looks like marketing optimism to me.

Now, on the general state of computer vision: If you're expecting terminator-like all-seeing computer in the near future, don't hold your breath! It might take some time:

At the moment, some object classes that don't vary too much in structure within the class (e.g., faces, cars, people), can be found reasonably quickly and moderately reliably in an image. To give an example, the detection of human faces in 800x600 images can be done in about a second, with about a 90-95% detection rate, but with about 1-10 false positive detections per image. That effectively means that if you find a face, there still is only a 30% change that it's actually a face.

In order to understand what you see, you rely on high-level semantics. These include the geometrical arrangement of objects (e.g., your head stands on top of your body, there is a hierarchy body->limbs->extremities, etc.) and general relations (e.g. finding faces at eye-level, so e.g. near the horizon). Research on these higher-level semantics is really in its infancy: the main problem is that it's very hard to get enough "world-knowledge" into the computer for it to make all the relations.

I can put a nice multiple-frame face-detection demo here, but that would destroy my research group's net-connection. If someone can offer a high-bandwidth spot, mail me: I'll then make a movie available.

Re:Bycatch is a big problem...

[tropicflite]

It was a tough call for me whether to mod you up, or add to your idea. Bycatch is a horrible problem that gets almost no notice. IMO, fishing in general is brutal and ugly, but I understand that not everyone is vegetarian (as I am). Anything that can be done to minimize the harm brought by the fishing industry to the the ocean environment (on the large scale) and to the individual sea creatures (on the small scale) is a step in the right direction.

If you eat fish you bear some of the responsibility for the bycatch problem by creating the demand. If the price of fish goes up a bit to pay for this equipment, that seems reasonable.

Not to go overboard (heh) on the topic, it's just responsible stewardship to minimize the negative impact of the fishing industry by fishing as cleanly and sustainably as possible.