YOLOv7 Fire Detection Model (DBA-YOLO)

This repository contains the implementation of the DBA-YOLO algorithm, a lightweight fire detection model based on YOLOv7-Tiny, enhanced with structural re-parameterization and multiple attention mechanisms. DBA-YOLO is designed for real-time fire and smoke detection in complex environments, achieving high precision and efficiency.

Overview of DBA-Fire Dataset

The DBA-Fire dataset is a comprehensive collection designed for fire and smoke detection in real-world scenarios. It consists of 3905 high-quality images with YOLO-format annotations. The dataset covers a wide variety of fire-related environments, including indoor and outdoor fires, ranging from small flames to large-scale fires in complex settings such as city roads, buildings, supermarkets, and parking lots.

Dataset Highlights

• Size: 3905 images
• Categories:
  • Flames
  • Smoke
  • Both flames and smoke
• Scenes: Diverse fire scenarios in both indoor and outdoor settings with varying complexities.
• Data Format:
  • Images: JPEG format (stored in DBA-Fire/JPEGImages/)
  • Labels: YOLO-format labels (stored in DBA-Fire/txt/), one label file per image.

Data Processing Pipeline

1. Data Collection: Images and videos of fire-related scenes were gathered using a web crawler, and video frames were extracted at a rate of 25 FPS.
2. Frame Segmentation: Frames were sampled every 3 frames to ensure variability.
3. Data Cleansing: Images were filtered based on several criteria such as object coverage and noise ratio.
4. Object Labeling: Images were manually annotated for flames and smoke using LabelImg, and converted to YOLO-format labels.

Intended Use

This dataset is primarily designed for training, fine-tuning, and evaluating fire and smoke detection models.

The DBA-Fire dataset will be available on Huggingface and GitHub after the acceptance of the associated paper. For more details, please visit DBA-YOLO Dataset.

---

DBA-YOLO Model Overview

DBA-YOLO is an enhanced version of YOLOv7-Tiny, incorporating attention mechanisms and structural re-parameterization for improved flame detection. The model introduces a new feature extraction module to boost performance and a focal modulation network to increase the receptive field of the detection layer.

Key Highlights:

• Average detection precision: 85.4%
• Mean Average Precision (mAP50): 85.6%
• Faster than YOLOv7 by 27% with an average detection time of 6.5ms per frame.
• Detection of fire, fire-like, and smoke-like objects with high precision and minimized false positives.

Abstract

In response to the challenges of detecting smoke-like objects in fire scenes and fire detection under various natural lighting conditions, we propose an improved fire detection algorithm based on YOLOv7-Tiny, named DBA-YOLO. This novel approach introduces an attention feature extraction module that combines structural re-parameterization, extending the feature extraction network architecture. It enhances feature propagation for flame target recognition and improves the performance of the network model. The integration of a focal modulation network into the head detection layer facilitates better feature exchange and enlarges the receptive field of the detection layer. By replacing the original model's detection head with a variable attention detection head, we significantly increase detection precision.

---

Training the Model

To train the DBA-YOLO model, follow the instructions below. The following configuration parameters are used for training:

• Epochs: 300
• Batch Size: 16
• Image Size: 640x640
• Optimizer: Adam
• Workers: 6

You can adjust the model path in train.py at line 529 to point to the correct location of your weights.

Training Command

Ensure that your environment is set up properly and dependencies are installed. Once ready, run the following command to start training:

python train.py --epochs 300 --img 640 --batch 16 --optimizer adam --workers 6 --data /path/to/dataset.yaml --weights /path/to/initial/weights.pt