### Table 2. OCR results for real-life blurred name card images after complete deblurring

"... In PAGE 5: ... However, OCR results are not satisfactory since ringing effects occur around most characters. We apply the blur estimation procedure in the second part of our method and obtain the following results in Table2 . Table 2.... ..."

### Table 8: Absolute Mean Deviation of the Restorations with respect to the Blurred Image and Non-blind Restoration

"... In PAGE 25: ... The authors believe that the smoother appearance of the non-blind restoration in Figure 16 (b) is due to the regularization method incorporated in the algorithm. The absolute mean deviation (AMD) for each of the restorations with respect to the blurred image is given in Table8 . The AMD is de ned as AMD( ^ f) = P8(x;y) jg(x; y) ? bf(x; y)j P8(x;y) jg(x; y)j where b = P8(x;y) g(x; y) P8(x;y) f(x; y) Table 8 gives the AMDs of the restored images of Jupiter.... In PAGE 25: ... The absolute mean deviation (AMD) for each of the restorations with respect to the blurred image is given in Table 8. The AMD is de ned as AMD( ^ f) = P8(x;y) jg(x; y) ? bf(x; y)j P8(x;y) jg(x; y)j where b = P8(x;y) g(x; y) P8(x;y) f(x; y) Table8 gives the AMDs of the restored images of Jupiter. The AMD provides, to some extent, a measure of the amount of \deblurring quot; by the corresponding restoration algorithm.... ..."

### Table 4: Relative Performance of Blurred Segmentation and Kernel Estimators in 2-D. The values shown are estimated values of E and from modeling the ratio of RMS errors for the Blurred segmentation and kernel estimators as E ( n n0 )? , with n0 = 10000 . Approximate standard errors are given in parentheses. Case Signal E

"... In PAGE 12: ... with n0 = 10000. Table4 shows the estimates values for E and obtained from this analysis. In all cases the values of are positive indicating that the blurred segmentation tends to increase its domination over the kernel method at larger sample sizes.... ..."

### Table 8: Absolute Mean Deviation of the Restorations with respect to the Blurred Image and

"... In PAGE 25: ... The authors believe that the smoother appearance of the non-blind restoration in Figure 16 (b) is due to the regularization method incorporated in the algorithm. The absolute mean deviation (AMD) for each of the restorations with respect to the blurred image is given in Table8 . The AMD is de ned as AMD( ^ f)= P 8(x;;y) jg(x;; y) ; bf(x;; y)j P 8(x;;y) jg(x;; y)j where b = P 8(x;;y) g(x;; y) P 8(x;;y) f(x;; y) Table 8 gives the AMDs of the restored images of Jupiter.... In PAGE 25: ... The absolute mean deviation (AMD) for each of the restorations with respect to the blurred image is given in Table 8. The AMD is de ned as AMD( ^ f)= P 8(x;;y) jg(x;; y) ; bf(x;; y)j P 8(x;;y) jg(x;; y)j where b = P 8(x;;y) g(x;; y) P 8(x;;y) f(x;; y) Table8 gives the AMDs of the restored images of Jupiter. The AMD provides, to some extent, a measure of the amount of \deblurring quot; by the corresponding restoration algorithm.... ..."

### Table 1. The PSF used to obtain the blurred image.

"... In PAGE 4: ...able 2. The adaptive filter coefficients at convergence. which it is clear that the method is useful in deblurring the degraded image. Table1 provides the PSF used. Ta- ble 2 shows coefficients of the adaptive filter at conver- gence.... ..."

Cited by 2

### Table 8 gives the AMDs of the restored images of Jupiter. The AMD provides, to some extent, a measure of the amount of \deblurring quot; by the corresponding restoration algorithm. The AMDs 25

"... In PAGE 25: ...Table8... In PAGE 40: ... Otherwise, go to 1. Table8 : Absolute Mean Deviation of the Restorations with respect to the Blurred Image and Non-blind Restoration Restoration AMD w.r.... ..."

### Table 3: Empirical Rates of Convergence in 2-D. The values shown are estimated values of from modeling the RMS error as C n? . Approximate standard errors of the apos;s are given in parentheses. Case Signal Blurred Segmentation Kernel Method

"... In PAGE 12: ... The power law characteristic is again seen to well describe the behavior of the RMS characteristic as a function of sample size. A summary of the estimated empirical rates of convergence are given in Table3 . The pattern is similar to what was seen in 1-D - rates of convergence achieved with blurred segmentation are better than those of the kernel method.... ..."

### Table 1: Empirical Rates of Convergence in 1-D. The values shown are estimated values of from modeling the RMS error as C n? . Approximate standard errors of the apos;s are given in parentheses. Case Signal Blurred Segmentation Kernel Method

### Table1. MSE of integrated and chromatic measures MSE non-blur

in AN IMPROVED METHOD FOR EXTRACTION OF INTRINSIC IMAGES FROM A SINGLE IMAGE WITH INTEGRATED MEASURES

"... In PAGE 5: ... In the estimated illumination image 2(d, h), the shadows are clearly shown. For comparison with our previous work, Table1 shows the Mean Square Error (MSE) of the estimated reflectance images and the ground truth - original synthetic object image, Fig. 2(i).... In PAGE 5: ... 2(i). In Table1 , the first row shows improved results obtained by the new method with integrated measures and the second row shows the results of the previous method with only chromatic measure. Note that the new method is rarely affected by blur shadows, and the previous one increases its error as the shadows get more blur.... ..."