QATM-KCNN: Improving Template Matching Performance based on integration of CNNs and with QATM by Kalman Filtering

Introduction: Processing large-scale aerial images from high-resolution cameras requires identifying numerous ground objects. However, detecting and recognizing these objects is time-consuming. Traditional template matching approaches have the potential to accelerate this process but suffer from sensitivity to various types of noise in high-definition images. Many template matching algorithms incorporate denoising techniques to mitigate this issue, and recent advancements in deep learning have significantly improved denoising methodologies. However, deep learning-based approaches require substantial labeled datasets, extensive computational resources, and deep expertise in data science. A fully trained deep learning model also demands intensive data processing over large-scale aerial imagery. Objectives: This research aims to develop a simple and effective pipeline that integrates deep learning and traditional computer vision methodologies. The objective is to enhance both speed and accuracy in template matching for high-resolution aerial images while minimizing computational costs and dataset requirements. Methods: The proposed approach applies a quality-aware template matching technique based on feature points with Kalman filtering. Additionally, a Convolutional Neural Network (CNN) model is employed to classify regions of interest using a pre-classified image dataset. After learning from these classified regions, the Kalman-filtered template matching model improves matching performance over high-resolution images. This hybrid approach leverages the strengths of both traditional template matching and deep learning, ensuring an efficient and scalable solution. Results: The proposed pipeline demonstrates significant improvements in both speed and accuracy compared to conventional template matching and standalone deep learning models. By integrating Kalman filtering with feature-based template matching and CNN classification, the system effectively reduces noise sensitivity while maintaining high detection accuracy. The results show enhanced template matching performance across various large-scale aerial image datasets. Conclusions: The developed approach offers an efficient and robust solution for template matching in high-resolution aerial imagery. By combining deep learning with traditional vision techniques, the pipeline optimizes computational efficiency and improves object detection accuracy. This methodology provides a practical and scalable alternative for preprocessing tasks in deep learning-based aerial image analysis.