A novel MPDENet model and efficient combined loss function for real-time pixel-level segmentation detection of tunnel lining cracks

Recent studies have demonstrated the potential of using convolutional neural networks (CNNs) for tunnel lining crack detection, offering a promising alternative to manual inspection methods. However, the performance of CNNs is often hindered by the sample imbalance in crack images, and the existing model architectures are often too computationally expensive, presenting challenges for real-time detection. In this study, an efficient combined loss function was proposed to address the sample imbalance issue, and a novel lightweight model, MPDENet, was introduced for real-time pixel-level segmentation detection of tunnel lining cracks of interference backgrounds. Specifically, the ResNet50 was replaced with improved MobileNetV2 as a backbone for the Pyramid Scene Parsing Network (PSPNet) to strike a desirable balance between detection accuracy and computational efficiency. Subsequently, the receptive field was expanded by including dilation convolution, facilitating the capture of more comprehensive crack contour information. Lastly, the ECA attention mechanism was introduced to enhance the model's capacity to distinguish cracks from interference backgrounds. Experimental results on a homemade dataset show that the combined loss function significantly improves the recall of crack detection and superior detection accuracy than other loss functions. MPDENet achieved an Intersection over Union (IoU) score of 75.26 % on the homemade dataset, which showcases a notable performance improvement of 2.64 % compared to PSPNet while significantly alleviating computational burden. Moreover, the robustness and scalability of the proposed method on two publicly available crack datasets are also verified. MPDENet can process 512 × 512 resolution crack images at a speed of 36 frames per second (FPS), making it well-suited for real-time detection in practical applications.