Research On Scene Recognition Based On Capsule Network

Among the various computer vision task solutions and tools,Convolutional Neural Network(CNN)is widely used in image classification,target recognition and other fields.As an image feature extractor,it can achieve good processing effects.However,CNN needs a large number of images to support training,and generally CNN uses pooling layer to conduct dimensioning reduction operation on features,which causes information loss and affects network expression ability.Capsule network(Caps Net)is a revolutionary neural network model in the field of computer vision in recent years.Different from CNN,Caps Net uses vector input and output to retain image spatial information,and uses dynamic routing mechanism to make the model have better feature fitting ability.However,in some specific scenarios and applications,Caps Net structure still needs to be optimized to achieve better results.This paper first introduces the background and significance of scene recognition research,and gives an in-depth understanding of the main methods used in scene recognition at home and abroad.Based on the principle of capsule network,the Scene Caps model suitable for scene recognition task is proposed.On the basis of the capsule network,convolutional layer＿2 is added,and this paper builts a five-layer Scene Caps network model,which increases the depth of the model and expands the dimensions in the capsule.The experiment analyzes the influence of learning rate of convolutive layer＿2 and convolution kernel size on recognition rate and training time.The experimental results of Scene15 datasets show that the highest recognition rate of Scene Caps model reaches 94.82%,with a time of 37.2h,which is 0.28% higher than that of capsule network model,and the training time is 11.8% lower.The paper analyzes the influence of changes in model parameters and training data volume on experimental results.The experimental results show that Scene Caps model can achieve the recognition rate of 75% of the capsule network data volume with 50% of the training data volume.Compared with the common scene recognition methods,the experimental results show that Scene Caps model has better recognition effect on scene images.Finally,an experiment is carried out on place2 datasets,and the experimental results verifies the effectiveness of the model.This paper proposes an optimization method based on 1*1 convolution kernel and batch normalization.The 1*1 convolution kernel is added to the convolutional layer＿2 to reduce the number of convolutional layer parameters and increase the nonlinear characteristics of the model.Experiments on the Scene15 datasets show that the model has a maximum recognition rate of 94.83% and training time of 32.2h,which is 0.01% higher than the recognition rate without the addition of 1*1convolution kernel and 13.4% lower than the training time.In order to maintain consistent network input distribution,batch normalization layer is applied in front of convolutional layer 1,convolutional layer＿2,and Primary Caps layer respectively,and then uses as input of activation function.By optimizing the Scene Caps model,the recognition rate increases from 94.82% to 95.03%,and the training time decreases from 37.2h to 20.1h.Experimental results verify the effectiveness of the optimization method.The main innovation points of this paper are that the scene recognition research method based on Scene Caps model and the optimization method based on 1*1convolution kernel and batch normalization are proposed for the first time,which effectively improves the scene image recognition accuracy and reduces the model training time.