Research On Object Recognition And Detection Algorithm Based On Deep Learning And Local Elastic Potential Energy Feature

In recent years,with the rapid development of mobile Internet and social networks,visual data including images and videos has grown rapidly.Extracting useful information from these visual data is still a challenge.In the past,people tried to extract and use the information through traditional machine learning algorithms,but they could not solve this problem effectively.The emergence of deep convolutional neural networks provides a method to effectively solve this problem.Based on deep learning algorithm and Local Elastic Potential Energy Feature,this paper deeply studies object recognition method and object detection method.The main work is as follows:1.We study Dense Net network?Res Net network and other classic neural networks,and propose residual module array based on Dense Net network and Res Net network to improve the recognition accuracy of existing neural networks.Two-dimensional convolutional network(TDNet),the basic structure of which is residual module array is designed in this paper and implemented with Tensorflow 1.4.0.TDNet is analyzed and compared with classic networks such as Res Net network in false recognition rate from the experimental point of view.Experimental results show that TDNet has a lower false recognition rate.At the same time,the influence of the depth,width,and growth number of residual module array on the recognition performance of TDNet is also studied from the experimental point of view.The simulation results show that increasing depth,width and growth number properly can effectively improve the recognition performance of TDNet.Additionally,the performance of TDNet in object detection tasks is also studied.Visualization results and experimental data show that TDNet can extract the features suitable for object detection tasks,which is superior to the classic object detection algorithms such as Faster RCNN,SPPnet,and SSD in terms of detection accuracy.2.In order to further improve the recognition accuracy of TDNet,We design the double path structure based on Dense Net network and grouped convolution method.The double path structure mainly includes the main path structure and the branch path structure.To illustrate its advantages,the neural network based on the double path structure(Double Path Dense Net,DPDNet)is compared with Res Net network,Dense Net network and other classic networks in the false recognition rate,parameter efficiency,calculation efficiency and other aspects of the comparative analysis.The simulation results show that DPDNet can achieve better recognition accuracy than Dense Net network with only a few more parameters.At the same time,the influence of growth number and fusion method on the recognition performance of DPDNet is also studied from the experimental point of view.The simulation results show that increasing growth number can effectively improve the performance of DPDNet,and different feature fusion methods will also affect DPDNet's performance.In addition,the anti-overfiting capability of DPDNet is analyzed.The analysis results show that DPDNet has good anti-overfitting capability.At the same time,we analyze whether the branch structure really plays a role in DPDNet through the average L1 norm.The simulation results show that the branch structure plays a role in DPDNet.Finally,the performance of DPDNet in object detection tasks is also analyzed.Visualization results and data analysis show that DPDNet can extract the features suitable for object detection tasks,which is better than the classical object detection algorithms.3.The commonly used traditional feature descriptors is studied in this paper.In order to improve the detection accuracy of traditional feature descriptors,we introduce Local Elastic Potential Energy Feature(LEPE),and analyze its properties theoretically.At the same time,a method for determining the elastic coefficient is proposed based on the convex optimization method.Besides,a cascaded classifier is designed,in which each base classifier uses five Ada Boost classifiers.These Ada Boost classifiers are trained based on Local Elastic Potential Energy Eigenvectors of different scales.Subsequently,the properties of LEPE are demonstrated and analyzed from the experimental point of view.The simulation results show that LEPE is suitable for object detection tasks.