DMASNet: A Defect Detection Network for PCBs via Multi-Scale Feature Collaborative Optimization

Technical Summary: DMASNet for PCB Defect Detection

Problem Statement
Reliable detection of defects on Printed Circuit Boards (PCBs) remains a significant challenge in modern manufacturing due to the wide variation in defect sizes and the complexity of manufacturing backgrounds. Traditional inspection methods face distinct limitations: manual visual inspection is labor-intensive and inconsistent; electrical testing risks physical damage to the board; and traditional Automated Optical Inspection (AOI) systems often struggle with small or morphologically irregular defects, requiring frequent parameter adjustments and lacking adaptability in flexible manufacturing environments. While deep learning offers a promising alternative, existing models often fail to balance the need for high-accuracy multi-scale feature extraction with the computational efficiency required for real-time industrial deployment.

Methodology
The authors propose DMASNet, a lightweight detection framework built upon the YOLO11 architecture, specifically optimized for PCB defect inspection. The network retains the standard three-stage structure (backbone, neck, and detection head) but introduces four key architectural modifications to address scale variation, background interference, and computational cost:

1. Dynamic Upsampling Reconstruction (DySample): To mitigate pixel aliasing and background interference common in static upsampling, the authors replace standard upsampling with the DySample module. This operator utilizes semantic information from deep, low-resolution feature maps to guide adaptive sampling points. By correcting spatial misalignment during feature upscaling, it enhances the representation of small defects while suppressing irrelevant background textures.
2. Multi-Scale Feature Extraction (MSCB): Recognizing that PCB defects vary significantly in pixel scale, the standard C3k2 module is enhanced with a Multi-Scale Convolutional Block (MSCB). This module integrates Multi-Scale Depthwise Convolution (MSDC) to capture complementary features across different spatial scales. It employs parallel depthwise convolutions to extract both local high-frequency textures and broader structural patterns, followed by channel shuffling to facilitate cross-scale feature interaction.
3. Lightweight Downsampling (ASDown): To offset the computational overhead introduced by the multi-scale feature extraction, the original single-stride convolution is replaced with the ASDown module. This module substitutes standard convolutions with depthwise separable convolutions, significantly reducing the parameter count and floating-point operations (FLOPs) while maintaining downsampling stability.
4. Bounding Box Regression Optimization (SIoU): The study adopts the SIoU (Scalable IoU) loss function to replace conventional IoU-based losses. SIoU incorporates angle, distance, and shape costs alongside the basic IoU term. This design introduces directional constraints and dynamic penalties based on angular deviation, aiming to improve localization accuracy for small and irregularly shaped defects and stabilize training convergence.

Key Contributions
The paper outlines five primary contributions:

• Upsampling Optimization: Introduction of a dynamic upsampling operator to correct spatial misalignment and reduce background interference without significant computational cost.
• Feature Extraction Optimization: Implementation of a multi-scale convolutional strategy to enhance the model's ability to detect defects of varying sizes under complex background conditions.
• Complexity Reduction: Deployment of the ASDown module to reduce overall model complexity, balancing the added cost of feature enhancement.
• Loss Function Optimization: Utilization of SIoU to improve localization accuracy and detection robustness, particularly for small targets.
• Experimental Validation and Deployment: Comprehensive evaluation on both a self-built SMT dataset and the public DeepPCB dataset, followed by successful deployment on an embedded platform (NVIDIA Jetson Orin Nano) to demonstrate real-time feasibility.

Experimental Results
The proposed model was evaluated against several state-of-the-art detectors (including RT-DETR-L and various YOLO versions) on two datasets:

• SMT Dataset (Self-Built): DMASNet improved the mean Average Precision at 50% IoU (mAP50) by 4.5 percentage points over the YOLO11n baseline. It achieved a detection speed increase of 20.4% while reducing the total parameter count by 0.41 million (from 2.58M to 2.17M). Notably, the ablation study indicated that the introduction of the ASDown module alone reduced the parameter count by 0.48M relative to the configuration immediately preceding its addition.
• DeepPCB Dataset (Public): The model demonstrated strong generalization, improving mAP50 by 0.6 percentage points over the baseline (98.9% vs 98.3%).
• Ablation Studies: The experiments confirmed that DySample and MSCB were the primary drivers for accuracy gains, while ASDown effectively reduced model size with a minimal trade-off in accuracy. SIoU further refined bounding box regression.
• Embedded Deployment: When deployed on an NVIDIA Jetson Orin Nano with TensorRT FP16 acceleration, DMASNet achieved an inference latency of 25.7 ms, supporting real-time inspection requirements in industrial environments.

Significance and Claims
The paper claims that DMASNet offers a favorable trade-off between detection accuracy and inference efficiency, making it particularly suitable for industrial PCB inspection where background complexity and defect scale variations are prevalent. By integrating dynamic sampling, multi-scale feature fusion, and lightweight downsampling, the model addresses the specific limitations of existing deep learning approaches in this domain. The successful deployment on an edge device demonstrates the practical engineering feasibility of the approach for real-time, on-line quality control. The authors note that the current system focuses on detection and recognition, with future work planned to integrate these outputs with downstream repair stations to create a closed-loop inspection-and-repair system.