SCS Undergraduate Thesis Topics
|Name||Academic Advisor||Thesis Topic|
|Fatma Tlili||Gianni Di Caro||Deep Learning and Pattern Analysis for Crack Detection|
To maintain infrastructures such as tunnels, bridges and buildings over time, efficient and accurate crack detection becomes a very important inspection step. Traditional methods for such investigation require experienced workers to manually examine these structures. These methods not only lack objectivity since they depend on the specialist’s knowledge, but are also very challenging and time-consuming due to the typical large surface area of the structures. A more recent trend in Computer Vision has led to the emergence of multiple approaches of automatic defect detection. These approaches range from image processing techniques (IPTs) such as Canny edges and Sobel filters which are used to segment the images and emphasize the edges of the cracks, to Machine Learning techniques (ML), which include neural networks techniques such as Multilayer Perceptrons (MLPs) and Convolutional Neural Networks (CNN). Nevertheless, the va- riety of the real-world situations such as the different degrees in shadowing, lighting and image resolutions can be very challenging for the image processing techniques especially in term of generalization capabilities. This is because the IPT require designing robust features that are good representative of the properties of the cracks, while also being flexible enough to adapt to the different real word situations. On the other hand, the machine learning techniques, and more specifically the deep learning techniques require large amounts of data and long training hours to achieve good performance. To over- come these limitations, we propose a crack detection system that combines both image processing as well as the use of CNNs to produce results that are expected to be more robust and generalizable to real-life situations even with limited available data. The motivation behind our system is its capability to avoid the challenges of hand crafting features that are specific to the cracks as it uses CNNs which are automatic feature extractors, and the features that they extract are well generalizable under different con- ditions. Our system is also not limited by the need of large amounts of data and long training hours because our pattern analysis methods alleviate the challenges as they improve the classification performance using spatial correlation techniques. The main steps of our system are preprocessing and decomposition of the images into small fixed size patches which can easily be classified through the CNN and can capture the structure of the cracks. This is followed by classifying each patch with a certain confidence by using a CNN. The CNN produces a probability map over the entire image of the likelihood of a crack being in the image. The next step is improving the probability map using spatial correlation techniques. This step is important because the probabilities of the CNN are independent from one patch to another, however these patches are spatially correlated since they are extracted from one common image which contains continuous longs cracks, therefore our pattern analysis techniques, exploits the proximity, the direction and the continuity properties in these cracks to revise the probability map produced by the CNN. The last step is stitching the patches to reconstruct the crack paths. The results show that combining the CNN with an image-processing component that exploits spatial correlation, improved the accuracy of the detection by 2.1% and the overall results of the system show 90.1% accuracy in detecting the cracks on concrete surfaces.