LDDP-Net: A Lightweight Neural Network with Dual Decoding Paths for Defect Segmentation of LED Chips

Jie Zhang, Ning Chen, Mengyuan Li, Yifan Zhang, Xinyu Suo, Rong Li, Jian LiuSensors 2025 · DOI

TL;DR

This paper introduces LDDP-Net, a lightweight neural network designed for precise defect segmentation on LED chips. By employing a modified MobileNetV3 backbone and a novel dual decoding path architecture, LDDP-Net effectively balances capturing global context with fine-grained detail recovery, achieving state-of-the-art accuracy with significantly fewer parameters and computational resources.

What Problem Does This Paper Address?

The semiconductor industry relies heavily on accurate defect detection to ensure the quality and performance of its products. For LED chips, which are fundamental components in modern lighting systems, identifying defects during manufacturing is critical. Traditional image processing methods often struggle with the complexity and variability of chip defects, leading to issues with efficiency, accuracy, and robustness, especially when dealing with large-scale production. While deep learning has shown promise, existing semantic segmentation models can be computationally expensive and may lose crucial fine-grained details during the upsampling process, hindering precise boundary localization of defects. This paper aims to bridge this gap by proposing a lightweight yet highly accurate model for LED chip defect segmentation.

Methods

The core of LDDP-Net is its novel architecture, which combines a modified encoder and a specialized decoder.

  • Modified Encoder: The authors adapt the MobileNetV3 backbone, known for its efficiency, by modifying its receptive field to prevent information loss, which is crucial for detailed segmentation. They also reduce the downsampling factor to retain more fine-grained features.
  • Dual Decoding Paths: The key innovation lies in the decoder, which features two parallel paths:
  • Coarse Decoding Path: This path uses a simple upsampling strategy (bilinear interpolation) to capture macro-level information, providing a general overview of the feature map.
  • Fine-Grained Decoding Path: This path is more intricate, employing multiple pooling and convolution techniques, along with deconvolution, to focus on recovering fine details and precise boundaries.
  • Multi-Layer Feature Fusion: To further enhance segmentation accuracy, intermediate-layer features from the encoder are integrated into the upsampling process. This fusion is guided by a Squeeze-and-Excitation (SE) block to adaptively modulate channel weights, ensuring that both semantic and detailed information are effectively combined.

Datasets

The authors utilized a custom dataset consisting of 150 defective LED chip images. These images were annotated into five categories: background, lead, electrode, light-emitting area, and defect. For evaluation, they employed five-fold cross-validation to ensure robust performance assessment.

Key Insights and Ideas

  • The Power of Dual Decoding: The paper highlights that a single upsampling strategy is insufficient for complex segmentation tasks. By employing both coarse (macro-level) and fine-grained (detail-level) decoding paths in parallel, the network can effectively reconstruct features at different scales.
  • Balancing Lightweight Design and Accuracy: The modification of MobileNetV3 demonstrates that efficient backbones can be adapted for demanding segmentation tasks without sacrificing significant performance, a crucial aspect for real-time industrial applications.
  • Intermediate Feature Fusion is Key: Integrating features from earlier layers of the encoder into the decoder's upsampling process is vital for accurate boundary localization, especially for irregularly shaped defects. The SE block's role in weighting these fused features is also significant.
  • Defect Segmentation is More Than Just Bounding Boxes: The paper implicitly argues for the importance of pixel-level segmentation over object detection for defects, as it allows for a more nuanced understanding and quantification of defect severity, which is often necessary for qualification.

Key Results

LDDP-Net achieved impressive results on the custom LED chip dataset:

  • High Accuracy: An mIoU (mean Intersection over Union) of 90.29% and an mPA (mean Pixel Accuracy) of 94.79%.
  • Lightweight Architecture: The model boasts a remarkably low parameter count of 2.98 million and 2.24 Giga FLOPs, making it significantly more efficient than many state-of-the-art methods.
  • Comparative Performance: LDDP-Net outperformed several established semantic segmentation models (FCN, PSPNet, EncNet, GCNet, DeepLabV3, SegNet) in terms of mIoU, often by substantial margins. For instance, it surpassed DeepLabV3 by 7.98%.
  • Robustness: Visualizations showed that LDDP-Net maintained better continuity of segmented targets and demonstrated superior robustness, particularly in handling complex defect patterns where other methods struggled.

Limitations and Future Work

  • Defect Quantification: While LDDP-Net excels at segmenting defects, the paper acknowledges that it does not quantitatively evaluate defects based on their types.
  • Future Directions: The authors suggest that future work should focus on extending the proposed network to enable the deployment and application of defect evaluation based on defect types, moving beyond simple segmentation to a more comprehensive analysis.

Summary updated

Abstract

Chip defect detection is a crucial aspect of the semiconductor production industry, given its significant impact on chip performance. This paper proposes a lightweight neural network with dual decoding paths for LED chip segmentation, named LDDP-Net. Within the LDDP-Net framework, the receptive field of the MobileNetv3 backbone is modified to mitigate information loss. In addition, dual decoding paths consisting of a coarse decoding path and a fine-grained decoding path in parallel are developed. Specifically, the former employs a straightforward upsampling approach, emphasizing macro information. The latter is more detail-oriented, using multiple pooling and convolution techniques to focus on fine-grained information after deconvolution. Moreover, the integration of intermediate-layer features into the upsampling operation enhances boundary segmentation. Experimental results demonstrate that LDDP-Net achieves an mIoU (mean Intersection over Union) of 90.29% on the chip dataset, with parameter numbers and FLOPs (Floating Point Operations) of 2.98 M and 2.24 G, respectively. Comparative analyses with advanced methods reveal varying degrees of improvement, affirming the effectiveness of the proposed method.
BibTeX
@inproceedings{zhang2025lddp,
  author = {Jie Zhang and Ning Chen and Mengyuan Li and Yifan Zhang and Xinyu Suo and Rong Li and Jian Liu},
  title = {{LDDP-Net: A Lightweight Neural Network with Dual Decoding Paths for Defect Segmentation of LED Chips}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25020425},
}
Figure