融合注意力机制与卷积神经网络的遥感图像微小目标检测方法

Remote Sensing Image Small Target Detection Method Combining Attention Mechanism and Convolutional Neural Network

摘要:

遥感图像由于几何畸变、传感器噪声、大气散射等多重因素的影响,存在细节模糊、辐射失真、空间错位等问题,使得整体质量下降,若是直接进行应用,会导致微小目标的检测性能下降。对此,提出融合注意力机制与卷积神经网络的遥感图像微小目标检测方法研究。通过几何校正、噪声抑制与质量增强步骤增强遥感图像质量。融合注意力机制与卷积神经网络构建微小目标检测模型,将增强后的遥感图像输入至训练好的微小目标检测模型中,获取微小目标分类结果(类别信息)与边界框回归结果(位置信息),从而实现了遥感图像微小目标的有效检测。实验结果显示:设计方法微小目标检测模型误检率最小值达到了0.25%；遥感图像峰值信噪比最大值达到了63dB,遥感图像融合特征提取误差最小值达到了0.2%,微小目标检测结果与实际结果高度吻合。

Abstract:

Due to multiple factors such as geometric distortion, sensor noise, and atmospheric scattering, remote sensing images suffer from problems such as blurred details, radiation distortion, and spatial misalignment, resulting in a decrease in overall quality. If directly applied, it will lead to a decline in the detection performance of small targets. A research on remote sensing image small object detection method integrating attention mechanism and convolutional neural network is proposed. Enhance the quality of remote sensing images through geometric correction, noise suppression, and quality enhancement steps. Integrating attention mechanism and convolutional neural network to construct a small object detection model, the enhanced remote sensing image is input into the trained small object detection model to obtain the classification results (category information) and bounding box regression results (position information) of small objects, thereby achieving effective detection of small objects in remote sensing images. The experimental results show that the minimum false detection rate of the small object detection model in the design method reaches 0.25%; The maximum peak signal-to-noise ratio of remote sensing images reached 63dB, and the minimum error of remote sensing image fusion feature extraction reached 0.2%. The results of small target detection are highly consistent with the actual results.