ADS Capstone Chronicles Revised

13

The model architecture uses a pre-trained MobileNet with ImageNet weights as the feature extractor. The last 20 layers of MobileNet are unfrozen for fine-tuning to adapt it for building damage classification. In addition, a custom convolutional branch with three convolutional blocks, including Conv2D layers, ReLU activation, batch normalization, and max pooling, extracts hierarchical features from the input images. The feature outputs from both branches are flattened and concatenated, then passed through two fully connected layers with ReLU activation, L2 regularization, and dropout for overfitting control. The final softmax output layer generates probabilities for the four damage categories. ResNet-50 which is a 50-layer deep convolutional neural network known for its residual connections and transfer learning capabilities with ImageNet pretrained weights, was employed in a two-stage model for building damage assessment in pre- and post-disaster images. In stage one, the model is trained for binary classification, grouping all damage categories into one to distinguish between affected and unaffected buildings. In stage two, the model fine-tunes on a balanced multi-class dataset, initialized with weights from stage one, leveraging transfer learning to improve classification performance across four damage categories. The model integrates a ResNet-50 backbone with custom convolutional layers for building damage classification. Additionally, input images are processed through the ResNet-50, which extracts global features, while custom convolutional layers capture fine-grained local details through convolution, batch normalization, ReLU activation, and max pooling operations. The outputs from these layers are flattened and concatenated to form a unified feature representation. Dense layers with L2 regularization and dropout are applied to reduce ResNet-50

4.5.2 Damage Classification Models

Simple CNN

CNNs are powerful frameworks for hierarchical feature learning, capable of extracting high-level features directly from raw image data through a data-driven approach. These features can be efficiently used for downstream tasks in an end-to end manner, making CNNs highly effective for a wide range of computer vision applications (Zheng et al., 2021). CNNs excel in image classification tasks due to their ability to autonomously capture spatial hierarchies of features from input images. The CNN model is structured with three convolutional blocks for hierarchical feature extraction. Each block includes a 3×3 convolutional layer with increasing filter sizes (32, 64, 128) and ReLU activation, followed by 2×2 max-pooling for downsampling and batch normalization to improve training stability and convergence. After the convolutional blocks, the model transitions to fully connected layers for classification. The extracted features are flattened and processed through a dense layer with 128 neurons and ReLU activation, along with a dropout layer (rate 0.5) to mitigate overfitting. Finally, a softmax-activated output layer produces class probabilities corresponding to the building damage categories. The MobileNet model is a lightweight deep learning architecture optimized for mobile and embedded vision tasks, making it ideal for resource-constrained environments. Its use of depthwise separable convolutions reduces computational load while retaining strong feature extraction capabilities. This project had fine-tuned a pre-trained MobileNet and added custom layers to effectively classify building damage while maintaining computational efficiency. MobileNet

279