Author Affiliations
School of Electronic Information, Xi’an Polytechnic University, Xi’an, Shaanxi 710048, Chinashow less
Fig. 1. Integral image schematic diagram
Fig. 2. An example of box filter. (a) Anisotropic directional derivative filter (θ=π/2); (b) corresponding box filter template and its size is 9×9
Fig. 3. Templates of box filter with six filter orientations. (a) θ=0; (b) θ=π/6; (c) θ=π/3; (d) θ=π/2; (e) θ=2π/3; (f) θ=5π/6
Fig. 4. Construction process of integral image
Fig. 5. Proposed corner detection algorithm
Fig. 6. Test image set for the evaluation about repeatability under affine transformation
Fig. 7. Average repeatability of the proposed method under different parameters. (a) Rotation; (b) shear transform
Fig. 8. Average repeatability under six image transformations for ten detectors. (a) JPEG compression; (b) rotation; (c) additive white Gaussian noises; (d) non-uniform scaling; (e) shear transform; (f) uniform scaling
Fig. 9. Test images and their ground truth corner positions. (a) Laboratory; (b) geometric image; (c) block image
Method | Addition operation | Multiplication operation | Division operation |
---|
Without rough screening | 5×M×N+(48+6)×M×N | (7×7+5)×M×N | M×N | Rough screening | 5×(M×N-Pn)+(48+6)×Pn | (7×7+5)×Pn | Pn |
|
Table 1. Influence of the coarse selection mechanism on the construction of corner measurement map
Method | Rotation | Non-uniform scaling | Shear | Uniform scaling | JPEG compression | Gaussian noise | Average |
---|
ANDD | 0.702 | 0.631 | 0.647 | 0.649 | 0.785 | 0.763 | 0.696 | CPDA | 0.794 | 0.773 | 0.784 | 0.767 | 0.809 | 0.832 | 0.793 | DOG | 0.772 | 0.710 | 0.684 | 0.794 | 0.734 | 0.695 | 0.732 | FAST | 0.795 | 0.771 | 0.780 | 0.769 | 0.811 | 0.822 | 0.791 | Method | Rotation | Non-uniform scaling | Shear | Uniform scaling | JPEG compression | Gaussian noise | Average | Harris | 0.793 | 0.675 | 0.731 | 0.686 | 0.725 | 0.608 | 0.703 | Harris-Laplace | 0.776 | 0.708 | 0.639 | 0.778 | 0.748 | 0.739 | 0.731 | KAZE | 0.751 | 0.791 | 0.674 | 0.809 | 0.701 | 0.673 | 0.733 | LF-Net | 0.565 | 0.675 | 0.648 | 0.640 | 0.852 | 0.790 | 0.695 | Zhang and Sun | 0.865 | 0.785 | 0.790 | 0.819 | 0.886 | 0.872 | 0.836 | Proposed | 0.862 | 0.814 | 0.820 | 0.831 | 0.901 | 0.881 | 0.852 |
|
Table 2. Average results for detectors under repeatability under affine transformation
Test image | Method | False corner number | Missed corner number | Localization error | Detected number |
---|
Laboratory | Zhang and Sun | 105 | 68 | 1.957 | 286 | Proposed | 70 | 90 | 1.885 | 229 | Geometric image | Zhang and Sun | 1 | 27 | 0.889 | 58 | Proposed | 6 | 25 | 1.255 | 65 | Block image | Zhang and Sun | 4 | 18 | 1.906 | 45 | Proposed | 0 | 16 | 2.040 | 43 |
|
Table 3. Detection performance of different algorithms
Method | Block(256 pixel×256 pixel) | Laboratory(512 pixel×512 pixel) | Boat(850 pixel×680 pixel) | Bike(1000 pixel×700 pixel) | Tree(700 pixel×1000 pixel) |
---|
Harris | 0.112 | 0.434 | 0.951 | 1.172 | 1.212 | Harris-Laplace | 0.124 | 0.451 | 1.614 | 1.874 | 1.945 | ANDD | 0.882 | 4.637 | 11.285 | 6.594 | 11.268 | Zhang and Sun | 2.184 | 8.362 | 23.735 | 35.469 | 36.289 | Proposed | 0.075 | 0.279 | 0.675 | 0.762 | 0.756 |
|
Table 4. Running time of different algorithms unit: s