Table 1.
Convolutional neural network related.
Table 2.
Traditional machine learning related.
Fig 1.
Forage grasses identification process.
Fig 2.
Photos of herbage species.
Fig 3.
Quantity before and after data balance for each type of forage grasses.
Fig 4.
E-A-Net network structure.
Fig 5.
Detailed structure of autoencoder network.
Fig 6.
Data processing results: The first row is the original data, the second is the cutout enhanced data, and the third is the background segmentation data.
Fig 7.
Loss curve of autoencoder network.
Fig 8.
Comparison of active thermograms before and after background removal.
Fig 9.
Training validation curve.
Table 3.
Train and validation results analysis.
Fig 10.
The influence of edge information on model recognition.
(a) Feature maps output by a Sobel operator, (b) Feature maps of the first three pooling layers of the network extracted for edges, (c) Extracting network feature map for class-activated edges.
Fig 11.
The influence of autoencoding network on the overall model.
(a) Image input to autoencoder network, (b) Result of last two convolutional layers and last pooling layer of autoencoder network, (c) Class activation E-A-Net autoencoder network partial characteristic diagram, (d) Class activation basic neural network thermogram.
Fig 12.
Confusion matrix diagram: Where the horizontal axis of the confusion matrix is the real type and the vertical axis is the prediction type.
On the first line is the confusion matrix for VGG16 [51], ResNet50 [38], and EfficientNetB0 [52] networks, and on the second line is the confusion matrix for E-A-VGG16 and E-A-ResNet50, E-A-EfficientNetB0 networks.
Table 4.
Test result.