基于YOLOv4的红外多波段图像目标检测算法

Infrared Multi-band Image Target Detection Algorithm Based on YOLOv4

不同波段的红外图像既具有信息差异性也具有相似性,每个波段与不同波段的组合也包含了丰富的语义信息。因此,在目标检测时充分利用不同波段图像的信息互补性,是提高目标检测和识别能力的有效途径。论文对红外多波段图像进行像素级融合,并提出了一种多个波段并联输入、并对单波段数据使用统一模型进行数据增强的方法,对图像信息进行表征,通过构建包含了同一场景下的多个波段信息的红外图像数据集,保证了不同波段信息数据增强的一致性。论文以YOLOv4网络模型为框架,利用红外多波段数据集进行单波段图像模型与多波段图像融合模型的精度对比实验,实验结果表明,多波段数据融合算法能够有效利用其子波段图像的正向信息,相较于单个红外波段的表现,mAP提升了10%以上,验证了该方法在多波段图像目标检测与识别方面具有优势。

Infrared images of different bands have both information differences and similarities,and the combination of each band and different bands also contains rich semantic information.Therefore,making full use of the information complementarity of images in different bands during target detection is an effective way to improve target detection and recognition capabilities.In this paper,pixel-level fusion of infrared multi-band images is proposed,and a method of parallel input of multiple bands and data enhancement using a unified model for single-band data is proposed to characterize the image information,and the consistency of data enhancement of different bands is ensured by constructing an infrared image dataset containing multiple band information in the same scene.Taking the YOLOv4 network model as the framework,this paper uses the infrared multi-band dataset to compare the accuracy of the single-band image model and the multi-band image fusion model.Experimental results show that the multi-band data fusion algorithm can effectively use the forward information of its sub-band images,and the mAP is improved by more than 10%compared with the performance of a single infrared band,which verifies that the proposed method has advantages in multi-band image target detection and recognition.