DLSS-YOLO: lightweight fuel tank flap detection model

Abstract

Automatic fuel tank flap detection represents a key technology for enabling unmanned refuelling stations. However, complex scenarios such as specular reflections, rain, and fog often directly affect the fuel tank flap detection. Conventional machine vision approaches suffer from low efficiency and insufficient accuracy. Moreover, existing lightweight detectors either sacrifice accuracy for compactness or remain too heavy for edge deployment. To address these limitations, an automatic fuel tank flap detection model named DLSS-YOLO (Dual-scale Lightweight Simple-ShuffleNet YOLO) is proposed based on an enhanced YOLOv8n. First, to achieve a lightweight backbone while maintaining detection accuracy, the ShuffleNetV2 network is introduced and extended with three novel modules, Simple-ShuffleUnit1&2 and DSWAM (Dual Spatial Weight Attention Mechanism), forming the proposed Simple-ShuffleNet backbone. Furthermore, to enhance feature fusion efficiency in the neck network, the model adopts a dual-scale detection head instead of large-target detection, reducing parameters by 33.9%. C2f-PStar module and MCAFF (Multi-Channel Adaptive Feature Fusion) module are proposed. Moreover, DSSC-Detect (Double Scale Shared Convolutional-Detect) head is proposed, integrating the DSWAM attention mechanism and shared convolutions for accurate recognition of small individual targets. Finally, to accelerate model convergence and improve regression accuracy, the Bifocal-SIoU loss function is introduced to optimize the bounding box regression loss, thereby effectively localizing fuel tank flaps. Experiments on a public dataset demonstrate that, compared to the original YOLOv8n, the proposed method reduces Parameters and giga floating-point operations per second (GFLOPs) by 74.8% and 60.5%, respectively, improves mean average precision @50% (mAP@0.5) to 99.4%, and yields a compact model size of only 1.8MB.