PRC-Light YOLO: An Efficient Lightweight Model for Fabric Defect Detection

PRC-Light YOLO: An Efficient Lightweight Model for Fabric Defect Detection

2024 | Baobao Liu, Heying Wang, Zifan Cao, Yu Wang, Lu Tao, Jingjing Yang, Kaibing Zhang
The paper presents the PRC-Light YOLO model, an efficient and lightweight approach for fabric defect detection. The model is an improved version of YOLOv7, designed to enhance detection accuracy while reducing computational load and model parameters. Key contributions include: 1. **Partial Convolution (PConv)**: Replaces 3×3 Conv operations in the Extended-Efficient Layer Aggregation Network (E-ELAN) to reduce computational complexity and memory access. 2. **Receptive Field Block (RFB)**: Introduces a multi-branch convolution structure to capture multiscale information and improve feature fusion. 3. **Content-Aware ReAssembly of Features (CARAFE)**: Replaces nearest neighbor interpolation with adaptive convolution kernels, enhancing semantic feature extraction. 4. **HardSwish Activation Function**: Replaces SiLU to reduce computational cost and improve inference speed. 5. **Wise-IOU v3 Bounding Box Loss**: Incorporates a dynamic non-monotonic focusing mechanism to mitigate adverse gradients from low-quality instances. Experiments show that PRC-Light YOLO reduces model parameters by 18.03% and computational load by 20.53%, while improving mAP by 7.6%. The model outperforms other object detection models in precision, recall, F1-score, and mAP, particularly in detecting Yard defects, Stains, and Holes. A fabric defect detection system is also designed to demonstrate the model's practical application. The study highlights the importance of dataset diversity and robustness in real-world scenarios, suggesting future research directions in these areas.The paper presents the PRC-Light YOLO model, an efficient and lightweight approach for fabric defect detection. The model is an improved version of YOLOv7, designed to enhance detection accuracy while reducing computational load and model parameters. Key contributions include: 1. **Partial Convolution (PConv)**: Replaces 3×3 Conv operations in the Extended-Efficient Layer Aggregation Network (E-ELAN) to reduce computational complexity and memory access. 2. **Receptive Field Block (RFB)**: Introduces a multi-branch convolution structure to capture multiscale information and improve feature fusion. 3. **Content-Aware ReAssembly of Features (CARAFE)**: Replaces nearest neighbor interpolation with adaptive convolution kernels, enhancing semantic feature extraction. 4. **HardSwish Activation Function**: Replaces SiLU to reduce computational cost and improve inference speed. 5. **Wise-IOU v3 Bounding Box Loss**: Incorporates a dynamic non-monotonic focusing mechanism to mitigate adverse gradients from low-quality instances. Experiments show that PRC-Light YOLO reduces model parameters by 18.03% and computational load by 20.53%, while improving mAP by 7.6%. The model outperforms other object detection models in precision, recall, F1-score, and mAP, particularly in detecting Yard defects, Stains, and Holes. A fabric defect detection system is also designed to demonstrate the model's practical application. The study highlights the importance of dataset diversity and robustness in real-world scenarios, suggesting future research directions in these areas.
Reach us at info@study.space
[slides and audio] PRC-Light YOLO%3A An Efficient Lightweight Model for Fabric Defect Detection