A hybrid approach to detect and classify pothole on Bangladeshi roads using deep learning

Potholes are a major problem for Bangladesh, a country with a developing economy and infrastructure. Traditional pothole detection methods, often based on manual inspection, are insufficient for effectively managing the vast national road network. The primary focus of this study is on pothole detection on roads in Bangladesh, employing deep learning techniques. The pothole dataset, comprising images captured by us, has been curated, leading to the development of two distinct datasets: one for pothole classification, totaling 6000 images, and another for pothole detection, comprising 1300 images. To improve the diversity of the datasets, the preprocessing procedure includes motion blur reduction and optical distortion correction. Our methodology employs state-of-the-art deep learning models, including Convolutional Neural Network (CNN), ResNet101, ResNet50, VGG16, DenseNet201, DenseNet169, InceptionResNetV2, MobileNetV2, EfficientNetB0, and a combined CNN along with ResNet101 model. The CNN and ResNet101 combination exhibit an exceptional accuracy of 97.89%. YoloV7 and YoloV8 demonstrate promising accuracy levels in pothole detection tasks, achieving 0.709 and 0.959, respectively. The datasets and models developed in this study have significant implications for improving road safety in Bangladesh. The proposed hybrid of CNN and pretrained ResNet101 exhibits the best performance than all other classification models applied in this study and the YoloV8 performed better than YoloV7 in the pothole detection task.