Multi-Scale Infrared Military Target Detection Based on 3X-FPN Feature Fusion Network

Abstract

To solve the problems of misdetection and omission of infrared military targets and poor detection effect in battlefield environments, an improved YOLOv4 algorithm is proposed to improve the accuracy of long-range target detection. First, a new 4th-scale feature extraction layer is introduced to enhance the multi-scale detection sensitivity for infrared military targets. Second, the TL intermediate layer channel is introduced to realize feature fusion across gradient connections, the 3X-FPN feature fusion network structure is proposed, and the adaptive network parameters are adopted to realize the weighted and balanced fusion of data to improve the target detection accuracy. Finally, the multi-scale detection loss function is established and optimized to improve the model stability and convergence effect. The depth separable convolution structure is adopted to realize the model’s lightweight. The experimental results of vehicle class military target ablation show that the improved algorithm increases the detection accuracy by 9.85% compared with the original algorithm, reduces the model volume by 36%, and its target detection distance is up to 2000 m. The improved algorithm achieves a mean average precision (mAP) value of 93.25% for multi-military target detection, which improves by 12.42% compared with the mainstream detection algorithm and meets the current combat data acquisition and processing requirements.

Keywords

• 3X-FPN feature fusion network
• long-range infrared military target detection
• multi-scale feature extraction
• multi-scale loss function optimization
• model lightweight