Machine learning optimization algorithm for real-time aerial optical detection and target recognition

Traditional ghost imaging technology faces a real-time bottleneck due to its reliance on multiple cumulative samplings and serial data processing. In recent years, optical ghost imaging has demonstrated potential in dynamic scenarios owing to its anti-interference capability and multi-dimensional information acquisition. However, its detection efficiency remains constrained by the substantial data processing requirements and the separation between target detection and recognition. To address these issues, a machine learning-integrated parallel processing framework for ghost imaging was proposed, which was integrated with an optimized YOLOv8 algorithm. By reconstructing an adaptive dynamic detection head module, incorporating multi-scale feature fusion, and employing a normalized Wasserstein distance loss function, the proposed approach enhances both detection precision and recognition accuracy in video streams (achieving nearly 0.7 mAP50-95 and over 95% precision). This framework enables real-time multi-dimensional processing from detection to recognition of small aerial targets. Experimental results demonstrate that the optimized algorithm achieves real-time inference in drone detection scenarios, offering a highly robust, end-to-end paradigm for multi-dimensional information processing in the field of optical engineering.

Optical ghost imaging (GI) technology has demonstrated significant potential for aerial target detection in dynamic scenarios due to its inherent advantages in anti-interference capability and multi-dimensional information acquisition. However, its practical application, particularly for real-time tasks, is severely hampered by two major bottlenecks: the reliance on multiple cumulative samplings for image reconstruction, which introduces latency, and the typical separation between the data-heavy imaging process and the subsequent target recognition task, which limits overall system efficiency. This paper aims to address these limitations by developing an integrated, high-performance framework that synergistically combines the physical principles of ghost imaging with an optimized deep learning model. The primary objective is to achieve highly accurate and real-time detection and recognition of small aerial targets, such as drones, directly from video streams generated under GI principles.

To overcome the stated challenges, a novel parallel processing framework is proposed, which integrates ghost imaging with an extensively optimized YOLOv8 algorithm. The core of our methodology lies in three key algorithmic innovations implemented within the YOLOv8 architecture. First, to enhance multi-scale feature representation crucial for small targets, a multi-level feature enhancement module (MLFEM) is embedded into the detection head. This module incorporates components like pyramid pool aggregation, a local attention gate (LAG), and adaptive spatial pooling (IMSAP), working in concert with reparameterized residual blocks (Rep Blocks) to dynamically weight and fuse cross-layer features, thereby strengthening detail perception and contextual information. Second, the standard detection head is replaced with a custom-designed asymmetric decoupled detection head (AsDDet). This head features decoupled branches for regression (REG) and classification (CLS), utilizing techniques like channel shuffling and depthwise convolutions to enhance spatial sensitivity and feature discriminability specifically for small or complex objects, while retaining the distribution focal loss (DFL) mechanism for precise bounding box decoding. Third, recognizing the inadequacy of traditional intersection over union (IoU) metrics for tiny objects, the loss function is improved by adopting the normalized wasserstein distance (NWD). The NWD loss models bounding boxes as Gaussian distributions and uses the Wasserstein distance for similarity measurement, providing robustness against minor localization deviations that disproportionately affect IoU, thus stabilizing training and improving performance for small targets.

The proposed optimized model was rigorously evaluated on a dedicated dataset of approximately 2000 images containing various small aerial targets like drones and birds, covering diverse and complex scenes. The experimental results demonstrate a substantial performance improvement over the baseline YOLOv8 model and other contemporary detectors. The key achievements include: a) A significant increase in mean average precision (mAP50-95) to 0.692 3, compared to 0.670 64 for the baseline YOLOv8, indicating superior overall detection quality across various IoU thresholds. b) A remarkable boost in precision, reaching a peak of 95.02%, which is approximately a 4.52% improvement over the baseline's 90.50%. This highlights the model's exceptional ability to minimize false positives, a critical requirement in practical surveillance and reconnaissance. c) Effective real-time inference capability, with an average processing latency of 41.57 ms per frame, far surpassing traditional imaging and detection systems like conventional video surveillance (100-500 ms), microwave radar (200-500 ms), and infrared systems (100-300 ms). d) Enhanced robustness in complex environments, as evidenced by the model's ability to correctly identify small drones against challenging backgrounds like clouds and foliage, where the baseline model produced false detections. Ablation studies further confirmed the synergistic contribution of the three core innovations (AsDDet, NWD, MlFEM), showing that their combined application yielded the best performance, whereas individual or pairwise integrations led to suboptimal or unstable results.

This research successfully presents a machine learning-optimized, parallel processing framework that effectively mitigates the real-time bottleneck of traditional ghost imaging for aerial target detection. By systematically enhancing the YOLOv8 algorithm through a reconstructed asymmetric decoupled head, a robust NWD-based loss function, and an advanced multi-scale feature fusion module, the proposed solution achieves a breakthrough in both detection accuracy and processing speed. The experimental validation confirms that the model not only sets a new benchmark in precision (95.02%) and mAP50-95 (0.692 3) for small target detection but also operates efficiently in real-time, fulfilling the stringent requirements of dynamic aerial scenarios. This work provides a highly robust, end-to-end paradigm for multi-dimensional information processing in optical engineering, with direct implications for military reconnaissance, airspace management, and security applications. Future work will focus on further model lightweight design for deployment on edge devices and exploration of cross-modal information fusion to further enhance capabilities.