Author Affiliations
1School of Electronic Information Engineering, Liaoning Technical University, Huludao, Liaoning 125105, China2Quanzhou Institute of Equipment Manufacturing Haixi Institutes, Chinese Academy of Sciences, Quanzhou, Fujian 362000, Chinashow less
Fig. 1. Structure of matrix capsule network
Fig. 2. Structure of wavelet capsule network
Fig. 3. KTH data set
Fig. 4. KTD data set
Fig. 5. UIUC data set
Fig. 6. Classification accuracies of DWTCapsNet on different texture data sets. (a) Train set; (b) test set
Fig. 7. Visualization of DWTCapsNet output feature
Fig. 8. Images after adding different levels of Gaussian noise
Fig. 9. Classification accuracies of wavelet capsule network at different noise levels (KTH dataset)
Layer | Output size | DWTCapsNet |
---|
Conv | 64×64 | kernel size: 7×7, stride: 2, with padding | DWT | 32×32 | wavelet decomposition | Dense block_1 | 32×32 | ×6 | Transition | 32×32 | kernel size: 1×1, stride: 1 | 16×16 | 2×2max pooling, stride: 2 | Dense block_2 | 16×16 | ×12 | Reduce Channel | 16×16 | kernel size: 3×3, stride: 1, with padding | PrimaryCaps | pose: 16×16×128 activation: 16×16×8 | kernel size: 1×1, stride: 1 | ConvCaps | pose: 7×7×256activation: 7×7×16 | kernel size: 3×3, stride: 2 | Class Capsule | pose: E×16activation: E | kernel size: 1×1, stride: 1 |
|
Table 1. Structural parameter of wavelet capsule network
Data set | KTH | KTD | UIUC |
---|
Train set | 3564 | 3136 | 700 | Test set | 1188 | 1344 | 300 | Class | 11 | 28 | 25 | Type | color | gray | gray | Total number of images | 4752 | 4480 | 1000 |
|
Table 2. Division of different data sets
Type | Model | KTH /% | KTD /% | UIUC /% | Reference |
---|
Traditional algorithm | LBP | 89.3 | 97.8 | 88.1 | Ref. [24] | ILBP | 96.9 | 99.7 | 91.9 | Ref. [25] | FLBP | 94.3 | 99.2 | 91.3 | Ref. [26] | GPBMFD | -- | -- | 90.30 | Ref. [27] | LQC | 96.39 | -- | 92.62 | Ref. [28] | SLGP | 95.60 | -- | -- | Ref. [29] | LCCMSP | 93.32 | 99.69 | -- | Ref. [30] | LQPAT | 83.76 | 96.92 | -- | Ref. [31] | CNN-based algorithm | T-CNN-3 | 99.4 | 73.2 | -- | Ref. [32] | VisualNet | 72.4 | 97.8 | -- | Ref. [33] | VGG-VD-16 | 97.8 | 99.5 | 93.3 | Ref. [11] | Deep-TEN | 84.5 | -- | -- | Ref. [34] | LFV | 82.6 | -- | -- | Ref. [35] | DWTCapsNet | 99.92 | 99.79 | 91.74 | ours |
|
Table 3. Classification accuracies of different texture classification algorithms
Data set | KTH-color | KTH-gray |
---|
Accuracy | 99.92 | 96.88 |
|
Table 4. Classification accuracies of DWTCapsNet on color and gray texture images unit: %
Data set | KTH | KTD | UIUC |
---|
DenseNet | 98.20 | 99.66 | 87.07 | DWTDenseNet | 99.38 | 99.72 | 87.83 |
|
Table 5. Classification result of DWTDenseNet unit: %
Data set | KTH | KTD | UIUC |
---|
MatrixCapsNet | 93.18 | 84.73 | 45.31 | DenseCapsNet | 93.91 | 99.01 | 88.12 |
|
Table 6. Classification result of DenseCapsNet unit: %
Rotation angle /(°) | Accuracy /% | Rotation angle /(°) | Accuracy /% | Rotation angle /(°) | Accuracy /% |
---|
0 | 99.71 | 120 | 41.26 | 240 | 41.26 | 30 | 67.25 | 150 | 57.10 | 270 | 50.99 | 60 | 44.60 | 180 | 96.66 | 300 | 43.11 | 90 | 51.27 | 210 | 64.27 | 330 | 62.00 |
|
Table 7. Classification accuracy of wavelet capsule network on rotating image data set