"... In PAGE 12: ... In all of these datasets, the number of numeric characters present is about 1/5 the number of alpha characters, so that combined results are heavily weighted by the performance on alphas. Table1 shows our character recognition results for the Phase I test at a 0% rejection level. The top of the table gives results when the field or character type is Unknown, while the bottom has the corresponding figures for when the type is Known.... In PAGE 12: ... The reason for this is that nearly all numerics have the potential to be confused with some alpha character; when these alternatives are removed, some of the ambiguous digits can be recovered. When we move to the right of the Table1 to look at the variable length hypothesis list results, we see a substantial improvement in truth containment rates, compared to that achieved using top choice. The overall inclusion for the Unknown condition is just short of 99%, with equivalent performance for digits and alphas.... ..."

"... In PAGE 12: ...8% decrease in the error rate. Ta- ble 1 and Table2 show the forced choice recognition rate on the the unmodified images and cropped image, respectively. Class Samples Cup Fighter Plane Car Bear Rabbit Cup 11 9 1 1 0 0 0 Fighter 10 0 10 0 0 0 0 Plane 13 0 0 13 0 0 0 Car 12 0 1 1 9 0 1 Bear 12 1 2 4 1 3 1 Rabbit 8 0 0 1 0 0 7 Table 1: Original forced-choice recognition confusion matrix.... ..."

"... In PAGE 27: ...index smpls 0 1 2 3 4 cup 0 288 282 0 6 0 0 ghter 1 144 0 120 7 16 1 snake 2 96 5 0 88 1 2 plane 3 144 0 2 7 135 0 sports car 4 120 1 0 6 1 112 Total hypotheses for class 288 122 114 153 115 Table1 0: Error matrix for generic classi cation experiment. Each row shows how the test images for a particular object class were classi ed.... ..."