MosQNet-SA is a novel deep learning model designed for accurate classification of mosquito species. This project focuses on distinguishing between three primary mosquito species: Anopheles, Aedes, and Culex, which are significant vectors for various diseases.
The dataset comprises 3,000 meticulously selected and annotated mosquito images:
- Anopheles: 1,000 images
- Aedes: 1,000 images
- Culex: 1,000 images
Sources:
The curated dataset used in this study is available on Kaggle: MosQNet-SA Dataset on Kaggle
These are the sample mosquito images
We conducted an extensive analysis of various pre-trained models:
- VGG16, VGG19
- ResNet50, ResNet101, ResNet152
- Xception
- InceptionV3, InceptionResNetV2
- MobileNet, MobileNetV2
- DenseNet121, DenseNet169, DenseNet201
- NASNetMobile
- EfficientNetB0, EfficientNetB1, EfficientNetB2
Our proposed model, MosQNet-SA, incorporates:
- Residual blocks
- Inception-like blocks
- MBConv blocks
- Spatial Attention mechanism
MosQNet-SA outperforms traditional transfer learning approaches:
| Model | Test Accuracy | Params | Trainable Params |
|---|---|---|---|
| MosQNet-SA | 99.42% | 388,349 | 384,155 (98.92%) |
| Metric | Anopheles | Aedes | Culex |
|---|---|---|---|
| Accuracy | 0.987752 | 0.998721 | 0.973467 |
| Precision | 0.993664 | 0.990643 | 0.979466 |
| Recall | 0.980208 | 0.99882 | 0.985537 |
| F1-score | 0.98689 | 0.994715 | 0.982492 |
We employed various augmentation techniques:
- Width and Height Shift (10%)
- Shear and Zoom Transformations (10%)
- Horizontal Flipping
- Nearest Neighbor Filling
- Optimizer: Adam
- Learning Rate: 0.001
- Epochs: 81
- EarlyStopping
- ModelCheckpoint
- TensorBoard
- ReduceLROnPlateau
- CSVLogger
- LearningRateScheduler
We utilized multiple XAI techniques for model interpretation:
- Saliency
- GradientInput
- GuidedBackprop
- IntegratedGradients
- SmoothGrad
- SquareGrad
- VarGrad
- GradCAM
- Occlusion
- RISE
- SobolAttributionMethod
- LIME
- KernelShap
MosQNet-SA demonstrates superior performance in mosquito species classification, achieving 99.42% accuracy with a significantly smaller model size compared to traditional transfer learning approaches.
- Expand the dataset to include more mosquito species
- Explore deployment on edge devices for real-time classification
- Investigate the model's performance in real-world scenarios
This project is licensed under the MIT License - see the LICENSE file for details.