Lightweight visible damage detection algorithm for embedded systems applied to pipeline automation equipment

Abstract

This research is designed for low-power, cost-effective and high-performance pipeline defect inspection in embedded systems. The backbone of the algorithm, CSPHet, employs efficient combinatorial convolution and heterogeneous kernel convolution, incorporates a lightweight convolutional structure SL in the neck of the network, enhances the nonlinear representation and feature processing capability through channel shuffling, utilizes the lightweight self-attention mechanism Detect_SA for prediction, and employs a multilayered GhostConv to improve the computational efficiency. In addition, the performance of the model is optimized by knowledge refinement. When tested on a customized pipeline defect dataset, CGYOLO used 25 % less memory, ran 39.5 % fewer gigaprocessors per second, had 28.5 % fewer parameters, and improved average accuracy by 4 % to 94.3 % compared to the smallest known state-of-the-art model. In addition, the algorithm demonstrates excellent lightweight performance and utility on the Kaggle concrete crack dataset and the Norbase dataset. Finally, the model has been successfully deployed in a real-time embedded system consisting of a Raspberry Pi 4B and low-cost embedded image sensors, as well as in a simulated pipeline interior environment, meeting the real-time requirements for inspecting visible pipeline damage.