CNN-Based Salient Target Detection Method of UAV Video Reconnaissance Image

In order to address the challenges of image complexity, capturing subtle information, fluctuating lighting, and dynamic background interference in drone video reconnaissance, this paper proposes a salient object detection method based on convolutional neural network (CNN). This method first preprocesses the drone video reconnaissance images to remove haze and improve image quality. Subsequently, the Faster R-CNN framework was utilized for detection, where in the Region Proposal Network (RPN) stage, the K-means clustering algorithm was used to generate optimized preset anchor boxes for specific datasets to enhance the accuracy of target candidate regions. The Fast R-CNN classification loss function is used to distinguish salient target regions in reconnaissance images, while the regression loss function precisely adjusts the target bounding boxes to ensure accurate detection of salient targets. In response to the potential failure of Faster R-CNN in extreme situations, this paper innovatively introduces a saliency screening strategy based on similarity analysis to finely screen superpixels, preliminarily locate target positions, and further optimize saliency object detection results. In addition, the use of saturation component enhancement and brightness component dual frequency coefficient enhancement techniques in the HSI color space significantly improves the visual effect of salient target images, enhancing image clarity while preserving the natural and soft colors, effectively improving the visual quality of detection results. The experimental results show that this method exhibits significant advantages of high accuracy and low false detection rate in salient object detection of unmanned aerial vehicle (UAV) video reconnaissance images. Especially in complex scenes, it can still stably and accurately identify targets, significantly improving detection performance.