The primary evaluation metrics obtained from training the CNN models using the Applied1 setting on the training data set. The models predict lensing probability values ranging between 0 and 1. However, to calculate the evaluation metrics, we have set a threshold of 0.99 to distinguish between lensed and non-lensed samples. The test data set comprises 96 072 samples evenly split between lensed and non-lensed categories.
| Model . | TP . | FP . |
|---|---|---|
| DenseNet121 | 0.899 | |$1.4 \times 10^{-3}$| |
| DenseNet169 | 0.914 | |$2.9 \times 10^{-3}$| |
| EfficientNetB3 | 0.946 | |$2.2 \times 10^{-3}$| |
| EfficientNetB4 | 0.951 | |$2.4 \times 10^{-3}$| |
| DenseNet121, EfficientNetB3 | 0.898 | |$6.8 \times 10^{-4}$| |
| DenseNet169, EfficientNetB3 | 0.900 | |$5.4 \times 10^{-4}$| |
| DenseNet121, EfficientNetB4 | 0.898 | |$6.2 \times 10^{-4}$| |
| DenseNet169, EfficientNetB4 | 0.910 | |$4.9 \times 10^{-4}$| |
| EfficientNetB3, EfficientNetB4 | 0.938 | |$9.5 \times 10^{-4}$| |
| DenseNet121, DenseNet169 | 0.886 | |$6.2 \times 10^{-4}$| |
| DenseNet121, EfficientNetB3, EfficientNetB4 | 0.897 | |$4.7 \times 10^{-4}$| |
| DenseNet169, EfficientNetB3, EfficientNetB4 | 0.906 | |$3.5 \times 10^{-4}$| |
| Model . | TP . | FP . |
|---|---|---|
| DenseNet121 | 0.899 | |$1.4 \times 10^{-3}$| |
| DenseNet169 | 0.914 | |$2.9 \times 10^{-3}$| |
| EfficientNetB3 | 0.946 | |$2.2 \times 10^{-3}$| |
| EfficientNetB4 | 0.951 | |$2.4 \times 10^{-3}$| |
| DenseNet121, EfficientNetB3 | 0.898 | |$6.8 \times 10^{-4}$| |
| DenseNet169, EfficientNetB3 | 0.900 | |$5.4 \times 10^{-4}$| |
| DenseNet121, EfficientNetB4 | 0.898 | |$6.2 \times 10^{-4}$| |
| DenseNet169, EfficientNetB4 | 0.910 | |$4.9 \times 10^{-4}$| |
| EfficientNetB3, EfficientNetB4 | 0.938 | |$9.5 \times 10^{-4}$| |
| DenseNet121, DenseNet169 | 0.886 | |$6.2 \times 10^{-4}$| |
| DenseNet121, EfficientNetB3, EfficientNetB4 | 0.897 | |$4.7 \times 10^{-4}$| |
| DenseNet169, EfficientNetB3, EfficientNetB4 | 0.906 | |$3.5 \times 10^{-4}$| |
The primary evaluation metrics obtained from training the CNN models using the Applied1 setting on the training data set. The models predict lensing probability values ranging between 0 and 1. However, to calculate the evaluation metrics, we have set a threshold of 0.99 to distinguish between lensed and non-lensed samples. The test data set comprises 96 072 samples evenly split between lensed and non-lensed categories.
| Model . | TP . | FP . |
|---|---|---|
| DenseNet121 | 0.899 | |$1.4 \times 10^{-3}$| |
| DenseNet169 | 0.914 | |$2.9 \times 10^{-3}$| |
| EfficientNetB3 | 0.946 | |$2.2 \times 10^{-3}$| |
| EfficientNetB4 | 0.951 | |$2.4 \times 10^{-3}$| |
| DenseNet121, EfficientNetB3 | 0.898 | |$6.8 \times 10^{-4}$| |
| DenseNet169, EfficientNetB3 | 0.900 | |$5.4 \times 10^{-4}$| |
| DenseNet121, EfficientNetB4 | 0.898 | |$6.2 \times 10^{-4}$| |
| DenseNet169, EfficientNetB4 | 0.910 | |$4.9 \times 10^{-4}$| |
| EfficientNetB3, EfficientNetB4 | 0.938 | |$9.5 \times 10^{-4}$| |
| DenseNet121, DenseNet169 | 0.886 | |$6.2 \times 10^{-4}$| |
| DenseNet121, EfficientNetB3, EfficientNetB4 | 0.897 | |$4.7 \times 10^{-4}$| |
| DenseNet169, EfficientNetB3, EfficientNetB4 | 0.906 | |$3.5 \times 10^{-4}$| |
| Model . | TP . | FP . |
|---|---|---|
| DenseNet121 | 0.899 | |$1.4 \times 10^{-3}$| |
| DenseNet169 | 0.914 | |$2.9 \times 10^{-3}$| |
| EfficientNetB3 | 0.946 | |$2.2 \times 10^{-3}$| |
| EfficientNetB4 | 0.951 | |$2.4 \times 10^{-3}$| |
| DenseNet121, EfficientNetB3 | 0.898 | |$6.8 \times 10^{-4}$| |
| DenseNet169, EfficientNetB3 | 0.900 | |$5.4 \times 10^{-4}$| |
| DenseNet121, EfficientNetB4 | 0.898 | |$6.2 \times 10^{-4}$| |
| DenseNet169, EfficientNetB4 | 0.910 | |$4.9 \times 10^{-4}$| |
| EfficientNetB3, EfficientNetB4 | 0.938 | |$9.5 \times 10^{-4}$| |
| DenseNet121, DenseNet169 | 0.886 | |$6.2 \times 10^{-4}$| |
| DenseNet121, EfficientNetB3, EfficientNetB4 | 0.897 | |$4.7 \times 10^{-4}$| |
| DenseNet169, EfficientNetB3, EfficientNetB4 | 0.906 | |$3.5 \times 10^{-4}$| |
This PDF is available to Subscribers Only
View Article Abstract & Purchase OptionsFor full access to this pdf, sign in to an existing account, or purchase an annual subscription.
