Results of the Fish Classification Project **new** 2003/12/16 20:18:19.718 US/Central 2003/12/16 20:19:39.164 US/Central Kyle Clarkson kclarks@rice.edu Jason Sedano jjsedano@rice.edu Ian Clark ianclark@rice.edu Kyle Clarkson kclarks@rice.edu Mark S. Husband mhusband@rice.edu Richard G. Baraniuk richb@rice.edu Jason Sedano jjsedano@rice.edu Ian Clark ianclark@rice.edu This shows the results for our fish classification project. The accuracy was good for both fish types with small amounts of noise, but it quickly decreased as noise was added.

General Results Two different sets of general tests were run. The first set was when an ideal picture is used and the fish is on a completely black background. The other set of tests that was run was for the situation when a non-ideal picture is used and the background on the picture must be subtracted.

Ideal Cases The ideal cases were extremely accurate with low levels of noise. With no noise, the classifier was able to correctly determine all the trout and 85% of the salmon. As noise was increase, the accuracy of the trout dropped off drastically, whereas the salmon continued to be detected more accurately as noise was added. Overall, this system works well and was extremely accurate.

Sample Pictures for the Ideal Tests Sockeye Salmon plain black background Steelhead Trout plain black background Tree Log with black background

Ideal Tests Accuracy Results Sockeye Salmon Steelhead Trout Trash/Unknown Items No Noise 85% 100% 66.67% Low Noise 85% 65% 20% Moderate Noise 85% 20% 60% Heavy Noise 50% 0% 80% Ideal Tests Confusion Results Sockeye Salmon Steelhead Trout Trash/Unknown Items No Noise 94.4% 87% 76.9% Low Noise 68% 86.7% 28.6% Moderate Noise 68% 100% 24% Heavy Noise 83% N/A 28.5%

Non-Ideal Cases The non-ideal cases occur when the image start with a set background, instead of having the fish placed on a black background. They are harder to detect because the picture of the fish is not completely accurate to begin with. Most of the test still work well, though, as the results below show.

Sample Pictures for Non-Ideal Tests Sockeye Salmon with background Steelhead Trout with background Tree Log with background

Non-Ideal Tests Accuracy Results Sockeye Salmon Steelhead Trout Trash/Unknown Items No Noise 46.7% 90% 10% Low Noise 80% 100% 0% Moderate Noise 73.3% 90% 10% Heavy Noise 73.3% 80% 0% Ideal Tests Confusion Results Sockeye Salmon Steelhead Trout Trash/Unknown Items No Noise 58.3% 47.4% 25% Low Noise 60% 66.7% N/A Moderate Noise 55% 64% 100% Heavy Noise 50% 61.5% N/A

The ideal cases performed basically as expected, with accuracy dropping as the noise was increased. As more and more noise was introduced into the system, more pictures were classified as unknown and fewer as fish. This is the way the system should behave and classify pictures which it is not sure of as unknown. The non-ideal cases actually showed an increase in accuracy as small noise was added to the pictures. Unlike the ideal cases, as noise was added, pictures became more likely to be classified as fish than they should be. The recognition of unknown pictures was very low and they were almost always classified as fish. This is not perfect behavior and if countries were to instal the system in a nonideal setting, they would want to improve the threshold values of the tests for their specific setting. All of the test worked very will with no noise in the system, but it was found that the accuracy of the fin detection test quickly decreased as noise was added to the system, and the length to width test began to fail once moderate noise was added. The intensity test and the feature detection tests continued to work well even through the high noise range, so they were definately the most reliable tests.