The Accelerated Inference of a Novel Optimized YOLOv5-LITE on Low-Power Devices for Railway Track Damage Detection

Abstract

Railway track malfunctions can lead to severe consequences such as train derailments and collisions. Traditional manual inspection methods suffer from inaccuracies and low efficiency. Contemporary deep learning-based detection techniques have challenges in model accuracy, inference speed, and are often associated with expensive computational costs and high power consumption when deployed on devices. We propose an optimized lightweight network based on YOLOv5-lite. which employs an enhanced Fused Mobile Inverted Bottleneck Convolution (BF_MBConv) to reduce the number of parameters and floating-point operations (FLOP) during backbone feature extraction. The Squeeze-and-Excitation (SE) mechanism is adopted, emphasizing more critical track features by assigning different weights from a channel-wise perspective. Utilizing DropBlock with holistic dropping as a substitute for Dropout with random dropping offers a more efficient means of discarding redundant features. In the neck section, Shuffle convolution replaces the conventional one, significantly reducing the parameter count while better integrating feature information post-group convolution. Lastly, the incorporation of Focal-EIoU Loss augments regression, and with the application of incremental dataset processing techniques, it addresses accuracy and sample imbalance issues. The refined algorithm achieves a mean Average Precision (mAP)@0.5 of 94.4%, marking an 8.13% improvement over the original YOLOv5-lite. Moreover, by leveraging the embedded platform integrated with the Intel ® Movidius™ Neural Compute Stick cluster as the portable device for model deployment, Achieved a frame rate of 18.7 FPS. Our findings indicate that this approach can efficiently and accurately detect railway track damages. Additionally, it addresses the previously overlooked issues of performance-cost trade-offs, countering the past trend of prioritizing high performance at the expense of elevated power consumption and costs, proposing a harmonized approach that prioritizes efficiency and affordability.

Keywords

• YOLOv5-lite
• lightweight network
• data enhancement
• neural compute stick cluster2
• energy-efficient consumption