Anomaly Detection in Water Distribution Assets using Spatial and Channel Attention based on DenseNet 201

The sector of water distribution management, anomaly detection isvery important for maintaining operational efficiency and reliability, especially in theregion where there are scarce resources. Though traditional CNNs (Convolutional Neural Networks) are effective in various image classification tasks, they often strugglewith recognizing complex patterns in real-world datasets due to their limited capacity to dynamically focus on relevant features. This thesis addresses these limitations by introducing an attention module on top of a CNN-based model to improve theperformance of the models. Results indicated that the attention-based DenseNet201 achieved an average recall of 95.53% and an F2 Score of 0.9552 on a novel dataset of tap images and it also outperformed the traditional CNN models. The enhancement of CNN with attention mechanism also highlighted the efficiency and accuracy of attention mechanisms in enabling the model to focus on important regions. This helped the model to work better even on images with complex backgrounds. This approach increases thecapabilities of CNN-based systems for anomaly detection. Also, it offers a better solutionfor the automated monitoring of water distribution assets. This eventually contributes to the reduction of water wastage and improving infrastructure maintenance