A Deeper CNN Approach For Detection Of Collapsed Buildings In Drone Images

Abstract

Building collapses lead to high mortality rate and other damages. UAVs attached with optical cameras are widely available and can be used for collapsed building detection as they can cover a larger area within a given time with minimal cost. This paper is a systematic study on utilization of deep learning based approaches for collapsed building detection on images collected by UAVs. This study aims to identify suitable deep learning models with focus on optimization for real world application scenarios. Five different pre-trained Convolutional Neural Networks (CNN) are trained using transfer learning on our custom data set to assess the performance. The real novelty in this study pertains to suitably enhancing the architecture of these models by increasing the depth and width of the neural network models on top of feature extraction and fine tuning approaches, without any substantial increase in the additional computational tasks. Another prominent contribution is the construction of a new drone-based building image data set with more emphasis on real world correlation between collapsed and normal buildings from disaster hit areas. The enhanced CNN architectures of the five models applied to this rich data set resulted in an average 6.67% increase in accuracy as compared to the general fine tuning approaches. Through deeper fine tuning approach, model training time was also reduced by 6 seconds in average with moderate increase in trained model size.