Research And Implementation Of Spatial Scene Target Detection Algorithm Based On YOLOv2

The algorithm based on deep convolutional neural network(DCNN)is more and more widely used and mature in social life.In the current increasingly tense international situation,the competition for space information power is particularly fierce.At present,it is more popular to use neural network to detect targets.It uses deep convolution kernel to extract target features,and finally obtains the position and category information of targets,so as to accurately locate and predict objects in the picture.Yolov2 algorithm is one of the representative works.Its detection accuracy is high,and it is especially suitable for local calculation of space object detection.However,on specific hardware equipment,the calculation amount of its space equipment is low,which is still a major computational bottleneck,which limits the performance of neural network in real-time detection on satellite equipment.Therefore,it needs to be adapted and improved accordingly.First,this paper collects sufficient data sets for satellite components,including more than 23000 pieces of data.At this data set processing level,a cutmix-m data enhancement method for satellite data sets is proposed.On the basis of yolov2,the backbone network is replaced,and the attention mechanism and feature fusion method are added.Through image processing optimization and algorithm optimization,the effect of important feature recognition is improved,and the computational burden of the model is reduced.The specific improvements are as follows.(1)Aiming at the problem of insufficient accuracy of yolov2 target detection algorithm,this paper proposes to improve and replace the original darknet-19 network backbone with the more advanced repvgg backbone network under mmdetection.Through the comparison of several network backbones through experiments,the advantages of the repvgg backbone are mainly reflected in high accuracy and small computation.(2)In view of the problem of limited data sets,a cutmix-m data enhancement method is redesigned,which preserves the original data labels by fusing four data set pictures into a new picture.After that,salt and pepper noise is added to generate a new dataset.The data set value is enhanced.(3)In view of the poor detection performance of some targets and serious interference from background noise,a method of adding ECA attention mechanism and feature fusion method to deep convolutional neural network is proposed,which strengthens the attention of important features,suppresses noise,improves the precision of individual classes,and enhances the detection performance of the overall model.The experimental results show that the yolov2-repvgg-e algorithm designed in this paper can achieve a comprehensive map value of more than 94.3% under the satellite data set collected and calibrated in this paper.Compared with other similar algorithms,the model calculation compression ratio of this paper ranks first.Compared with the official darknet-19 neural network,the accuracy of the algorithm in this paper is 91.2%,and the calculation amount can be reduced to 76.48% of the official calculation amount.On the basis of the experimental results,the size of the network model can be compressed by 17.82%.