The Research On Object Recognition And User Attributes Analysis Based On Instantiation Parameter Compressed Capsule Networks

With the further development of image and text data processing to the deep learning of the intelligent processing,the machine learning algorithm based on the deep neural networks such as convolution neural networks,recurrent neural networks,generative adversarial networks have emerged.As a new intelligent processing algorithm,capsule networks has become a hot topic in the related research fields.Capsule networks is a novel neural networks structure,which can improve classification performance with small data sample set.Compared with the traditional convolution neural networks,it has better classification performance but more artificial neurons of feature,which leads to the poor calculated performance.By analyzing the difference between the capsule networks and the traditional convolution neural networks,it can conclude that the model compression method suitable for the traditional neural networks will no longer suitable for the capsule networks.This impedes the development of the capsule networks in the direction of deep stack.Aiming at the problems of capsule networks model compression,data reduction and the performance optimization of the Object Identification and user attribute analysis tasks.The thesis concludes the following researches:1.Facing the calculated performance problem of capsule networks,the energy efficiency analysis of the capsule networks is presented,and the energy consumption heat zone of the capsule networks is obtained in its primary capsule layer.According to the results,an IPC-CapsNet compression algorithm is proposed based on the structural characteristics of the capsule networks.The algorithm can reduce the computational complexity and compress the scale of the model computation on the basis of maintaining the accuracy of model classification.The experiments are carried out on standard dataset MNIST,Fashion-MNIST and UMIST(a face recognition dataset),and the following experimental results are obtained: The compression ratio of 9.4%,the top-1 accuracy rate of 97.01% and the acceleration rate of 40% were obtained on the MNIST dataset.The compression rate of 6.3%,the top-1 accuracy rate of 86.31% and the acceleration rate of 36% were obtained on the Fashion-MNIST data set.The compression ratio of 6.3%,the top-1 accuracy rate of 80.21% and the acceleration rate of 47% were obtained on the UMIST dataset.The experimental results show that the algorithm can achieve better compression rate and acceleration rate within acceptable loss of model accuracy.2.In order to solve the problem of dimensional reduction which is caused by the curse of dimensionality,the thesis put forwards the method of Dimension reduction(FSLLE)based on fast K select algorithm,by using data reduced sample data set by FSLLE algorithm will speeded up convergence performance in model training stage.Combined with IPC-CapsNet and face recognition experiment,the following experimental results are obtained: By comparing with PCA,ISOMAP,LLE,LTSA,it is concluded that FSLLE achieves the best results and greatly improves the computational performance compared with the LLE algorithm.Combined with IPC-CapsNet,the reduced dimensions data of the FSLLE algorithm obtained 76.41% accuracy after 100 rounds of training,leading the LLE algorithm 22.1th.The experimental results show that the algorithm achieves a good balance between dimensions reduction and time consumption.Meanwhile,in face recognition experiments,the convergence performance of the model is improved by using the dimensions reduction FSLLE algorithm data to train.The vehicle types varied greatly and high requirements of the running speed exists as a problem in the vehicle identification tasks in the road monitoring.Based on the algorithm of IPC-CapsNet and FSLLE,the thesis proposes the object identification algorithm(MS-CapsNet-R-CNN)based on the multi-scale capsule networks which can improve the performance of vehicle identification under the condition of satisfying the running speed.The experiments are carried out on standard data sets PASCAL VOC 2007,PASCAL VOC 2012 and vehicle Identification data set BIT-Vehicle.The following research works have been done: The average recognition rate on PASCAL VOC 2007 is 55%,which leads R-CNN algorithm 15.29% and Faster RCNN algorithm 0.73% respectively.The average recognition rate on PASCAL VOC 2012 is 48.98%,which leads R-CNN algorithm 13.99% and Faster RCNN algorithm 0.75% respectively.On the BIT-Vehicle data set,the average recognition rate of 74.56% is obtained by using multi-scale pretreatment,which is 7.15% higher than that without multi-scale preprocessing.The experimental results show that the multi-scale networks model is superior to the single-scale networks model in vehicle recognition and has a better average recognition rate.Aiming at the user attribute classification problem of the user profiling extraction of Weibo data,this dissertation analyzes the characteristics of Weibo data,and label the emotion labels for other related data set by transfer learning.Then combined with IPC-CapsNet,a user gender classification algorithm is proposed.It has conducted a gender classification experiment based on sentiment analysis on Weibo data set,and obtained the following experimental results: Weibo data is tested by combining transfer learning with SPT-CapsNet,the accuracy of gender classification was raised from 84.20% to 89.73%,and increased by 5.53%.At the same time,the form of centralized feature engineering is compared and the LSTM layer has the best feature extraction capability.At the same time,the form of centralized feature engineering is compared and the LSTM layer has the best feature extraction capability.In the tasks of object identification and user attribute analysis,the thesis proves that the method of the IPC-CapsNet and Dimension reduction methods(FSLLE)can compress the model and process the problem of dimension reduction while ensuring the accuracy.Good results can be obtained in the corresponding experiments: it improves the operation speed,maintains the balance of performance and efficiency and expands the application scope of the CapsNet.