DACA-Net: Detail-aware network with contrast attention for locating liquid crystal display defects

Introduction

This repo contains the official PyTorch implementation of DACA-Net.

  Result on LCD light defect dataset          Result on LCD surface defect dataset

Highlight

• Proposed a novel LCD defect detection network to improve the detection of tiny defects and low-contrast defects.
• Refined CSPDarknet53 network to enhance detail awareness for tiny objects.
• Developed a low-level semantic deep fusion module to improving the detection performance for tiny objects.
• Proposed a dual-focus contrast enhancement attention module to improve low-contrast object detection.
• Made the first dataset for LCD defect detection available.

Installation

Our codebase is based on YOLOv5. You only need to follow its instructions for installation.

Dataset Preparation

LCD light defect dataset

The LCD light defect dataset consists of images displayed on a 7-inch screen with a resolution of 768×1280 pixels, covering spot, line, and mura defects. It includes 1,608 images in total. Due to the limited number of actual defects available during production, data augmentation was applied, resulting in 225 spot, 483 line, and 900 mura. The dataset is provided by the data fusion research team at the University of Electronic Science and Technology of China. To download the dataset, please visit: https://pan.baidu.com/s/1R7OENxkxPrY5RVweAtToPg?pwd=1357

Samples

LCD surface defect dataset

The surface defect dataset included three types of defects: oil, scratches and stains, with 400 images per defect type at a resolution of 1920×1080. The dataset is built and presented by Jian Zhang, Miaoju Ban (Open Lab on Human Robot Interaction, Peking University). To download the dataset, please visit: https://robotics.pkusz.edu.cn/resources/dataset/.

Samples

oil                   scratch                   stain

PCB surface defect dataset

The PCB defect dataset contained 693 images with six types of defects: missing holes, open circuit, mouse bites, spur, short, and spurious copper. The dataset is built and presented by Lihui Dai et al. (Open Lab on Human Robot Interaction, Peking University). To download the dataset, please visit: https://robotics.pkusz.edu.cn/resources/dataset/.

Samples

Result

Result on LCDLD dataset

| Models | P (M) | P (L) | P (S) | R (M) | R (L) | R (S) | AP (M) | AP (L) | AP (S) | mAP | Params | FLOPs | |-----------|-------|-------|-------|-------|-------|-------|--------|--------|--------|------|--------|-------| | YOLOv5s | 99.8 | 89.9 | 94.7 | 100 | 83.3 | 95.0 | 99.5 | 85.6 | 94.9 | 93.3 | 7.0 | 16.0 | | YOLOXs | 99.4 | 86.8 | 89.7 | 100 | 82.2 | 86.3 | 99.5 | 94.4 | 90.2 | 91.4 | 10.6 | 23.6 | | YOLOv6s | 99.5 | 66.2 | 94.4 | 100 | 87.8 | 79.3 | 99.5 | 86.4 | 91.9 | 92.6 | 16.3 | 44.2 | | YOLOv7 | 98.9 | 91.9 | 92.7 | 100 | 88.9 | 93.3 | 99.5 | 88.2 | 93.8 | 93.8 | 37.2 | 105.2 | | YOLOv8s | 99.2 | 78.9 | 95.1 | 100 | 77.8 | 87.1 | 99.5 | 84.0 | 95.0 | 92.8 | 11.1 | 28.6 | | YOLOv9s | 99.9 | 83.4 | 95.8 | 100 | 82.2 | 86.8 | 99.5 | 85.3 | 94.9 | 93.3 | 9.6 | 38.7 | | YOLOv10s | 98.8 | 90.6 | 87.9 | 97.2 | 84.4 | 90.8 | 99.4 | 91.1 | 92.5 | 94.3 | 8.0 | 24.5 | | YOLOv11s | 99.8 | 91.5 | 94.3 | 100 | 83.3 | 90.0 | 99.5 | 89.5 | 94.2 | 94.4 | 9.4 | 21.3 | | Ours | 100.0 | 93.0 | 92.8 | 100 | 93.2 | 96.2 | 99.5 | 94.8 | 95.7 | 96.7 | 7.4 | 20.3 |

The model weight files can be downloaded at: https://pan.baidu.com/s/1ECJpvRn4xe-UCIrBAuGTCg?pwd=1357.

Result on PKU-Market-Phone dataset

| Models | P (O) | P (SC) | P (ST) | R (O) | R (SC) | R (ST) | AP (O) | AP (SC) | AP (ST) | mAP | |--------------|-------|--------|--------|-------|--------|--------|--------|---------|---------|------| | YOLOv5s | 98.3 | 96.4 | 97.0 | 98.8 | 95.6 | 97.2 | 98.6 | 96.7 | 96.2 | 97.2 | | YOLOXs | 98.4 | 96.8 | 97.1 | 97.0 | 87.3 | 93.4 | 98.9 | 96.2 | 96.4 | 96.2 | | YOLOv6s | 97.1 | 94.0 | 95.3 | 97.6 | 95.8 | 88.1 | 98.9 | 97.0 | 94.3 | 96.7 | | YOLOv7 | 97.7 | 96.5 | 98.3 | 98.8 | 96.3 | 97.6 | 98.8 | 96.3 | 97.6 | 97.6 | | YOLOv8s | 97.6 | 91.3 | 94.7 | 99.2 | 97.1 | 89.1 | 99.1 | 97.8 | 94.6 | 97.2 | | YOLOv9s | 97.2 | 96.3 | 94.6 | 98.2 | 95.0 | 86.5 | 98.9 | 97.1 | 95.5 | 97.2 | | YOLOv10s | 93.9 | 93.7 | 95.1 | 96.4 | 94.4 | 85.6 | 98.0 | 96.9 | 94.6 | 96.5 | | YOLOv11s | 95.8 | 92.7 | 92.7 | 98.8 | 96.8 | 91.8 | 99.2 | 97.6 | 94.9 | 97.2 | | Ours | 99.4 | 95.0 | 97.8 | 98.8 | 95.8 | 98.3 | 99.3 | 97.6 | 98.4 | 98.5 |

The model weight files can be downloaded at: https://pan.baidu.com/s/1dDz-8PBUB9IYvf89bs_g?pwd=1357.

Result on PKU-Market-PCB datasett

| Metrics | YOLOv5s | YOLOXs | YOLOv6s | YOLOv7 | YOLOv8s | YOLOv9s | YOLOv10s | YOLOv11s | Ours | |--------- |---------|--------|---------|--------|---------|---------|----------|----------|------| | P (Mh) | 98.8 | 98.4 | 98.4 | 91.7 | 99.1 | 99.1 | 96.5 | 97.7 | 98.5 | | P (Mb) | 91.5 | 95.9 | 82.5 | 82.1 | 93.7 | 94.0 | 96.7 | 93.8 | 92.8 | | P (Oc) | 95.4 | 95.9 | 92.0 | 93.5 | 95.0 | 96.3 | 94.8 | 95.4 | 97.3 | | P (Sh) | 97.4 | 98.2 | 95.4 | 96.5 | 94.8 | 95.1 | 95.8 | 95.7 | 96.1 | | P (Sp) | 96.3 | 96.3 | 85.6 | 94.1 | 98.2 | 95.1 | 97.7 | 95.1 | 95.8 | | P (Sc) | 91.2 | 93.8 | 83.9 | 96.2 | 97.5 | 97.3 | 93.1 | 89.0 | 98.0 | | R (Mh) | 99.1 | 99.1 | 98.2 | 98.9 | 99.5 | 98.6 | 97.3 | 99.1 | 99.1 | | R (Mb) | 90.4 | 78.3 | 78.3 | 83.1 | 86.7 | 91.3 | 94.4 | 86.7 | 96.4 | | R (Oc) | 98.1 | 80.0 | 81.0 | 84.6 | 89.5 | 92.0 | 92.9 | 97.7 | 100 | | R (Sh) | 96.6 | 96.0 | 86.2 | 93.8 | 95.0 | 97.4 | 91.4 | 95.5 | 97.4 | | R (Sp) | 81.4 | 76.0 | 82.7 | 73.5 | 71.6 | 83.3 | 84.3 | 76.3 | 85.3 | | R (Sc) | 96.0 | 79.2 | 80.2 | 82.2 | 89.1 | 87.9 | 92.1 | 89.1 | 97.2 | | AP (Mh) | 99.3 | 99.1 | 98.9 | 98.7 | 99.4 | 98.8 | 99.0 | 99.0 | 98.8 | | AP (Mb) | 89.8 | 90.8 | 83.1 | 87.6 | 92.4 | 95.3 | 93.4 | 92.7 | 96.5 | | AP (Oc) | 98.5 | 85.9 | 90.8 | 90.9 | 95.2 | 94.8 | 95.7 | 99.1 | 99.4 | | AP (Sh) | 99.2 | 97.8 | 94.5 | 95.2 | 97.6 | 98.5 | 96.3 | 98.0 | 97.8 | | AP (Sp) | 86.4 | 80.9 | 77.7 | 78.6 | 88.7 | 87.3 | 85.6 | 86.7 | 85.3 | | AP (Sc) | 97.1 | 88.8 | 85.2 | 92.1 | 93.1 | 98.0 | 94.1 | 92.4 | 98.2 | | mAP | 95.2 | 90.6 | 88.3 | 90.5 | 94.4 | 95.4 | 94.0 | 94.7 | 97.1 |

The model weight files can be downloaded at: https://pan.baidu.com/s/1zj2D1yZ1SHY-j2yJOWyEZg?pwd=1357.

Acknowledge

The code base is built with ultralytics. Thanks for the great implementations!

Citation

Please cite the following paper if the code and dataset help your project:

bibtex @article{luo2024daca, title={DACA-Net: Detail-aware network with contrast attention for locating liquid crystal display defects}, author={Luo, Sijie and Chen, Huaixin and Liu, Biyuan}, journal={Displays}, pages={102913}, year={2024}, publisher={Elsevier} }