Author Affiliations
School of Electronics Information, Xi'an Polytechnic University, Xi'an, Shaanxi 710048, Chinashow less
Fig. 1. AlexNet model structure
Fig. 2. Activation function
Fig. 3. Dataset samples. (a)(b) Black; (c)(d) yellow; (e)(f) green; (g)(h) brown; (i)(j) blackish green; (k)(l) orange; (m)(n) blue; (o)(p) purple
Fig. 4. Neural network training process
Fig. 5. Improved neural network model structure
Fig. 6. Impact of training sample number on test accuracy
Fig. 7. Convolution feature map. AlexNet model (a) layer 1 feature map, (b) layer 3 feature map, (c) layer 5 feature map; improved model (d) layer 1 feature map, (e) layer 3 feature map, (f) layer 5 feature map
Fig. 8. Accuracy and loss curves of AlexNet model. (a) Training and test accuracy; (b) loss curve
Fig. 9. Accuracy and loss curves of improved model. (a) Training and test accuracy; (b) loss curve
Network | Layer | Conv1 | Pooling1 | Conv2 | Conv3 | Pooling2 | Conv4 | Conv5 | Conv6 | Pooling3 |
---|
I | 8 | 3×3 | 3×3 | 3×3 | 3×3 | 3×3 | 3×3 | 3×3 | — | 3×3 | II | 8 | 5×5 | 3×3 | 5×5 | 5×5 | 3×3 | 5×5 | 5×5 | — | 3×3 | III | 8 | 3×3 | 2×2 | 3×3 | 3×3 | 2×2 | 3×3 | 3×3 | — | 2×2 | IV | 9 | 3×3 | 3×3 | 3×3 | 3×3 | 3×3 | 3×3 | 3×3 | 3×3 | 3×3 |
|
Table 1. Four kinds of network structures based on AlexNet
Network | I | II | III | IV |
---|
Training time /h | 2.4 | 2.1 | 2.7 | 3.1 | Accuracy rate /% | 89.31 | 81.74 | 85.20 | 88.42 |
|
Table 2. Four kinds of network performance
Layer | Layerinput | Convolution kernel | Convolutionoutput | Pooling | Layer output |
---|
| Size | | Number | Step | Size | Step |
---|
Conv1 | 127×127×3 | 3×3 | 64 | 1 | 127×127×64 | — | — | 127×127×64 | Pooling1 | 127×127×64 | — | — | — | — | 3×3 | 2 | 63×63×64 | Conv2 | 63×63×64 | 3×3 | 128 | 1 | 63×63×128 | — | — | 63×63×128 | Conv3 | 63×63×128 | 3×3 | 128 | 1 | 63×63×128 | — | — | 63×63×128 | Pooling2 | 63×63×128 | — | — | — | — | 3×3 | 2 | 31×31×128 | Conv4 | 31×31×128 | 3×3 | 256 | 1 | 31×31×256 | | | 31×31×256 | Conv5 | 31×31×256 | 3×3 | 256 | 1 | 31×31×256 | — | — | 31×31×256 | Pooling3 | 31×31×256 | — | — | — | — | 3×3 | 2 | 15×15×256 | FC1 | 15×15×256 | — | — | — | — | — | — | 256 | FC2 | 256 | — | — | — | — | — | — | 8 |
|
Table 3. Improved AlexNet model parameters
Model | AlexNet | Improved model without technique | Improved model with technique |
---|
Amount of error | 56 | 43 | 28 | Error rate | 0.373 | 0.286 | 0.186 |
|
Table 4. Impact of optimization techniques on model
Algorithm | Training time /h | Training accuracy /% | Test accuracy /% | Number of parameters /M |
---|
Algorithm of Ref. [4] | 0.2 | 81.6 | 74.3 | — | Algorithm of Ref. [7] | 3.3 | 95.3 | 82.6 | 25.2 | AlexNet | 5.4 | 98.1 | 61.2 | 60.3 | VGG-16 | 6.3 | 99.2 | 67.7 | 138.1 | Improved algorithm | 2.4 | 91.5 | 88.2 | 15.8 |
|
Table 5. Comparison of recognition performance of different methods