Robust CNN-Based Segmentation of Infrastructure Cracks Segregating from Shadows and Lines

Detecting cracks early in various structures is crucial not only for preserving the integrity of these structures but also for safeguarding lives, especially in environments like coal mines, bridges, and buildings. However, employing deep learning(DL) and computer vision for crack identification in mines poses significant challenges. These challenges arise from the presence of intensity inhomogeneity in cracks and the intricate nature of the background such as elements, shadows, lines, and water-dripping patterns, which can resemble more like cracks. Moreover, these techniques often struggle to detect extremely small cracks, further complicating the detection process. Manually inspecting these extensive structures is highly arduous. Utilizing DL techniques in conjunction with image processing can prove immensely beneficial in the early identification of these cracks. In this study, we improve the performance of TernausNet with modified loss functions. The encoder component of TernausNet is substituted with various networks, including VGG-16 [19], Resnet34 [7], and Resnet101 [7]. Additionally, the study investigates the utilization of pre-trained weights from ImageNet [16] for the encoder component and compares it against training from random initialization for crack detection using TernausNet.