Pixel-level detection of multiple pavement distresses and surface design features with ShuttleNetV2

Abstract

Concurrently detecting multiple objects of interest will yield massive time savings in processing and enable a more streamlined and unified detection system. The ShuttleNet is designed to repeat the encoding–decoding round freely or even endlessly, achieving prodigious successes in terms of simultaneous detection of multiple pavement distresses and surface design features on asphalt pavements. This paper proposes an efficient and improved architecture of ShuttleNet called ShuttleNetV2 for enhanced global modeling and retrieving fine details capabilities. The proposed ShuttleNetV2 represents two major modifications on the original ShuttleNet. On the one hand, the self-attention mechanism is purposefully introduced to capture long-range dependency. On the other hand, ShuttleNetV2 adopts various sampling scales to combine the characteristics of different receptive fields. The experimental results indicate that the recommended architectural variation of the proposed ShuttleNetV2 model yields a mean F-measure of 94.21% and a mean intersection-over-union of 0.8914 on 1500 pairs of testing images. The proposed ShuttleNetV2 outperforms ShuttleNet in detecting nearly all types of pavement patterns. In particular, ShuttleNetV2 efficaciously tackles the tangible limitations of ShuttleNet in detecting giant distresses. Moreover, the ShuttleNetV2 can process an image in roughly 78 ms using modern graphic processing unit devices, which has a promising potential in supporting the real-time detection of multiple pavement distresses and surface design features on asphalt pavements.