my portfolio

Post-earthquake road debris detection based on deep neural networks and post-event very high-resolution images

My thesis delves into the refinement of post-earthquake road debris detection using deep neural networks and high-resolution images. Initiating the process with Python Colab, VHR images underwent meticulous division into 256x256 dimensions, coupled with noise reduction using specialized libraries. Subsequently, the labelme library was employed to produce masks and labels, paving the way for the implementation of five conventional neural networks. Faced with the limitations of these networks in detecting road debris, attention module, and dense skip connections were strategically introduced into the deep learning architecture. The Unet network emerged as the frontrunner, showcasing superior performance among its counterparts. Further elevating the Unet model, the integration of SE Attention (Squeeze and Excitation Attention) and concurrent use of dense skip connections resulted in a noteworthy 6% improvement in IoU, which metric is important to us. This enhanced and modified network not only exhibited superior IoU metrics but also demonstrated heightened accuracy in detecting road debris. Evaluation metrics encompassed Loss Function, Accuracy, IoU, and Recall.