Fig. 1. (a) Original image and (b) energy distribution after DCT
Fig. 2. Structure of CNN
Fig. 3. Three-dimensional spectrum diagram. (a) DCT coefficient spectrum of original image; (b) spectrum after the coefficient selection
Fig. 4. Structure of DCT-CNN model
Fig. 5. Flow chart of landform image classification algorithm based on DCT and deep network
Fig. 6. Example images in database. (a) UC Merced LU database; (b) UAV landing landform database
Fig. 7. Classification performance of each method when the number of training samples is different. (a) UC Merced LU database; (b) UAV landing landform database
Fig. 8. Image classification confusion matrix for UAV landing landform database
Layer | Type | Patch size | Stride | Zero padding | Output size |
---|
x | Input | | | | 128×128 | h1 | Convolution | 5×5 | 1 | 2 | 128×128×32 | h2 | ReLU | | | | | h3 | Mean pooling | 3×3 | 2 | | 64×64 | h4 | Convolution | 3×3 | 2 | 0 | 32×32×32 | h5 | ReLU | | | | | h6 | Max pooling | 3×3 | 2 | | 16×16 | h7 | Convolution | 7×7 | 1 | 2 | 14×14×64 | h8 | ReLU | | | | | h9 | Max pooling | 3×3 | 2 | | 7×7 | h10 | Convolution | 7×7 | 1 | 0 | 1×1×64 | h11 | ReLU | | | | | h12 | Convolution | 1×1 | 1 | 0 | 1×1×10 | o | SVM | | | | n(class) |
|
Table 1. Layer parameters of DCT-CNN network structure
Method | Accuracy /% | SD | Training time /h |
---|
Method 1 | 84.25 | 0.78 | 0.8 | Method 2 | 95.76 | 0.28 | 1.0 | Method 3 | 92.83 | 0.52 | 3.3 |
|
Table 2. Effect of different methods on classification of UC Merced LU database
Method | Accuracy/% | SD | Training time /h |
---|
Method 1 | 83.73 | 0.85 | 1.0 | Method 2 | 94.38 | 0.34 | 1.3 | Method 3 | 92.10 | 0.61 | 3.9 |
|
Table 3. Effect of different methods on classification of UVA landing landform database
Method | Accuracy /% | SD |
---|
RF | 79.25 | 0.82 | LDA-RF | 82.92 | 0.69 | CS-CNN[5] | 92.86 | 0.59 | PSR[15] | 89.10 | 0.69 | MS-DCNN[16] | 91.34 | 0.63 | DCT-CNN | 95.76 | 0.28 |
|
Table 4. Comparison of the classification accuracy of different methods for UC Merced LU database
Method | Accuracy /% | SD |
---|
RF | 77.10 | 0.70 | LDA-RF | 80.23 | 0.74 | CS-CNN[5] | 91.78 | 0.62 | DCT-SAE[12] | 86.49 | 0.96 | MS-DCNN[16] | 90.16 | 0.71 | DCT-CNN | 94.38 | 0.34 |
|
Table 5. Comparison of the classification accuracy of different methods for UAV landing landform database