From: Comparing pre-trained models for efficient leaf disease detection: a study on custom CNN
Model name | Architecture | Parameters (millions) | Image input size | Top-1 accuracy (%) | Top-5 accuracy (%) | Notable features |
---|---|---|---|---|---|---|
Densenet201 | Dense connectivity | 20.2 | 224 × 224 | 76.6 | 93.3 | High parameter efficiency, feature reuse |
EfficientNetB3 | Efficient architecture | 12.2 | 300 × 300 | 82.2 | 96.1 | Compound scaling for better efficiency |
EfficientNetB4 | Efficient architecture | 19.3 | 380 × 380 | 83.5 | 96.7 | Improved depth and width for larger models |
Inception ResnetV2 | Inception + ResNet | 55.9 | 299 × 299 | 80.4 | 95.3 | Multi-level feature extraction, residual connections |
MobilenetV2 | Mobile-friendly | 3.4 | 224 × 224 | 71.8 | 91.0 | Depthwise separable convolutions, lightweight |
ResNet152 | Residual connections | 60.4 | 224 × 224 | 77.8 | 93.8 | Deeper version of ResNet50, more representational power |
Resnet50 | Residual connections | 25.6 | 224 × 224 | 76.1 | 92.9 | Identity shortcuts, widely used architecture |
Vgg16 | Simplicity and depth | 138.4 | 224 × 224 | 71.6 | 90.0 | Classic architecture with multiple convolutional layers |
Xception | Depthwise separable convs | 22.9 | 299 × 299 | 79.0 | 94.5 | Similar to InceptionV3 but with depthwise convolutions |