Intelligent Defect Detection of Design Materials Based on Edge AI

Abstract

Aiming at the problems such as the imbalance between accuracy and efficiency, insufficient adaptability of multi-scale defects and limited inference resources in the deployment of design material defect detection at the edge end, this paper proposes a lightweight detection method based on edge AI. In the training stage, multi-branch topology is used to enhance feature expression, and in the inference stage, multi-branch convolution is merged into a single standard convolution by algebraic fusion. Under the condition that the number of parameters is only 2.2M and the amount of computation is 2.6 GFLOPs, the detection accuracy is significantly improved. The bidirectional feature pyramid and lightweight channel-spatial joint attention fusion module are designed to effectively cope with the challenges of defect scale diversity, low contrast and inter-class similarity, and the recall rate of small-scale defect reaches 82.4%. An early departure inference mechanism based on input complexity and a dynamic computing path guided by confidence are constructed, and combined with quantization-aware training and operator fusion optimization, 18ms end-to-end inference latency and 55.6 FPS frame rate are achieved on Jetson Xavier NX, with power consumption of only 2.4 W. Experiments on NEU-DET, DAGM2007 and self-built industrial datasets show that the proposed method mAP@0.5 reaches 75.2%, which outperforms the mainstream lightweight detection network with more than three times the number of parameters, and shows superior performance in a variety of defect types and scales. It provides a feasible technical path for real-time, safe and low-cost detection in intelligent manufacturing scenarios.