基于YOLOV3的改进目标检测识别算法

Improved Target Detection and Recognition Algorithm Based on YOLOV3

经过近几十年不断的研究和发展,红外目标检测识别在侦察、导弹制导等领域取得了卓越的成就和广泛的应用,亦成为当今的热门话题。为进一步提高模型的检测识别性能,提出一种基于YOLOV3改进的目标检测识别算法。首先,通过分析红外目标的检测特性,改进了原始算法的特征提取网络,融合KL-LOSS,在原网络预测目标位置的基础上,进一步预测了位置的准确度标准差,并结合Soft-NMS算法用于改善网络的检测准确度;其次,针对红外目标相对三通道彩色图像的特征量少的问题,在检测层前融合了SKNET模块,使网络更加关注目标的有用特征;最后,给出改进网络训练的新的损失函数及前向传播算法流程。实验结果表明:改进的KS-YOLO网络在目标域(实拍空中红外目标数据集)上的平均AP性能值要优于原来的YOLOV3网络2.4个百分点,预测时间比YOLOV3实用性更好、更快。

After decades of continuous research and development,the technology of infrared target detection and recognition has been widely applied in the fields such as military reconnaissance and missile guidance,and has made remarkable achievements.Therefore,the technology has become a hot topic today.In order to improve the detection and recognition performance of the technology,an improved target detection and recognition algorithm based on YOLOV3 is proposed in this paper.First,by analyzing the detection characteristics of infrared targets,the feature extraction network of the original algorithm is improved,and the KL-LOSS function is integrated.The target location is predicted based on the original network,based on which the accuracy standard deviation of the location is further predicted,and the detection accuracy of the network is improved in combination with the Soft-NMS algorithm.Then,in order to solve the problem that the number of feature quantities of an infrared target is less than that of a threechannel color image,the SKNET module is integrated before the detection layer so that the network could pay more attention to the useful target features.Finally,a new loss function and the forward propagation algorithm process of the improved network training are presented.The experimental results show that the average precision(AP)performance of the improved KS-YOLO network in the target domain(real-shot aerial infrared target dataset)is 2.4%better than the original YOLOV3 network,and the prediction time is better and faster.