HOGnet-driven Siamese architecture for quantifying complexity of infrared countermeasure scenarios

• Siamese Network for Infrared Complexity: Proposes a Siamese network-based method to quantify the complexity of infrared scenes, enhancing countermeasure effectiveness evaluation for infrared-guided missiles. • HOGnet and NCC Integration: Introduces HOGnet, a model combining traditional HOG features with deep learning, and integrates a normalized cross-correlation (NCC) layer for improved robustness against environmental variations. • New Complexity Metrics: Develops three task-oriented metrics-CDGB, LDTI, and IDLB-to quantify scene attributes affecting recognition accuracy in infrared countermeasure scenarios. • Strong Correlation with Detection Performance: Extensive experiments show that the proposed metrics exhibit strong inverse correlations with detection accuracy, demonstrating their effectiveness in performance evaluation. Infrared scene complexity serves as a critical foundation for evaluating the countermeasure effectiveness of infrared-guided missiles in challenging environments. Conventional complexity metrics are constrained by the limited expressive power of traditional image features, making them insufficient for evaluating and optimizing deep learning-based systems. To overcome these limitations, the present study introduces an innovative approach to infrared scene complexity assessment, utilizing a Siamese network architecture. Initially, the HOG feature extraction mechanism is encoded as a neural network model, termed HOGnet, which merges interpretability with the learning capacity of neural networks and is adaptively trained to deliver discriminative visual representations. Furthermore, a normalized cross-correlation (NCC) layer is incorporated to enhance robustness against illumination variation and background clutter. Based on similarity response maps generated by the Siamese architecture, three metrics, Confusion Degree of Global Background (CDGB), Loss Degree of Target Information (LDTI), and Influence Degree of Local Background (IDLB), are derived to quantify scene attributes such as occlusion and background clutter that influence recognition accuracy. Extensive experiments show that the proposed complexity metrics exhibit strong inverse correlations with detection performance (0.9130 with center error and - 0.9238 with IoU), which has positive significance for performance evaluation and enhancement of infrared imaging systems.