基于深度学习的高分辨率图像的智能检测

针对高分辨率图像下目标所占面积小,检测效果较低、实时性较差的问题.提出了一种基于LDCF-ResNet50的深度学习模型检测方法.以行人检测为例说明此方法的有效性.首先基于局部无关通道特征的预测提议区域,对行人潜在区域粗检测.然后,设计候选区域合并和扩展方法,将合并后的区域用于后面的ResNet-50神经网络.其次,设计了一个合适的ResNet-50网络,用于精确检测该区域.最后将ResNet-50网络的检测结果映射到原始图像中,输出检测结果.为了验证本文所提方法的有效性,在清华-戴姆勒数据库平台上对高分辨率图像进行实验验证.实验结果表明,所提出方法能够有效地检测行人.与主流的算法(包括Faster R-CNN,YOLOv3和SSD)相比,本文方法对行人检测的平均精度分别提高了4.07％、17.79％和31.45％.

Intelligent Detection for High-resolution Images Based on Deep Learning

Aiming at the problems of small target area, low detection effect and poor real-time performance under high-resolution images. This paper proposes a deep learning model detection method based on LDCF-ResNet50. This article uses pedestrian detection as an example to illustrate the effectiveness of this method. Firstly, the proposed area is predicted based on LDCF (Locally Irrelevant Channel Features), and the pedestrian potential area is roughly detected. Then, design candidate region merging and expansion methods, and use the merged region in the following ResNet-50 neural network. Secondly, a suitable ResNet-50 network is designed to accurately detect the area. Finally, the detection result of the ResNet-50 network is mapped to the original image, and the detection result is output. In order to verify the effectiveness of the LDCF-ResNet50 method proposed in this paper, the high-resolution images are experimentally verified on the Tsinghua-Daimler database platform. Experimental results show that the proposed method can effectively detect pedestrians. Compared with mainstream algorithms (including Faster R-CNN, YOLOv3 and SSD), the average accuracy of pedestrian detection by this method is increased by 4.07%, 17.79% and 31.45%, respectively.