Enhanced Water Puddle Segmentation and Detection Using DCU-Net

In recent years, the development of artificial intelligence (AI) and the expanding applications of deep learning, particularly in image segmentation and recognition using convolutional neural networks (CNNs), have been remarkable. With climate change posing increasing threats to road safety through hazards, such as potholes, water puddles, and cracks, accurately detecting and segmenting these dangers has become crucial for protecting road users. Detecting and segmenting water puddles is particularly challenging due to their varying sizes and shapes. This article proposes a novel dense connection U-net (DCU-Net) architecture specifically designed to improve the detection accuracy of water puddles. We modify and combine elements of DenseNet-121 and U-Net for model training, with DenseNet-121 handling encoding tasks and U-Net managing decoding tasks. A bridging layer is employed to enhance feature transmission between the encoder and decoder stages, effectively recovering more image details. We compared the performance of DCU-Net with Deeplabv3+, U-Net, PSPNet, and ResU-Net on segmentation tasks using our data set. DCU-Net achieved an intersection over union (IoU) score of 82%, along with the best accuracy and F1-score. These results demonstrate the effectiveness and potential of DCU-Net in detecting water puddles in diverse environments, providing a new architecture for future research in related fields.