ReSTiNet: An Efficient Deep Learning Approach to Improve Human Detection Accuracy.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Human detection is crucial for security and safety monitoring.
• Deep learning has advanced human detection, but challenges remain in creating small, fast, and accurate models.
• Existing compact networks like SqueezeNet and MobileNet face limitations in balancing size, speed, and accuracy.

Purpose of the Study:

• To introduce ReSTiNet, a novel small convolutional neural network for human detection.
• To address the limitations of existing models regarding network size, detection speed, and accuracy.
• To optimize network architecture for minimal parameters and enhanced feature propagation.

Main Methods:

• Developed ReSTiNet, a compact convolutional neural network utilizing fire modules with optimized residual blocks.
• Evaluated the number and position of fire modules to minimize model parameters and network size.
• Compressed the Tiny-YOLO architecture, enhancing feature propagation and information flow.

Main Results:

• ReSTiNet achieved a model size of 10.7 MB, five times smaller than Tiny-YOLO.
• Demonstrated superior performance with mAP of 27.3% on MS COCO and 63.74% on PASCAL VOC datasets.
• Outperformed SqueezeNet and MobileNet in terms of mean Average Precision (mAP) on benchmark datasets.

Conclusions:

• ReSTiNet offers a highly efficient solution for human detection, balancing small model size, fast detection speed, and high accuracy.
• The proposed network architecture effectively resolves overfitting issues.
• ReSTiNet provides a significant advancement in compact deep learning models for real-world computer vision applications.