Research On Image Recognition Algorithm Based On Lightweight Convolutional Neural Network

With the fast development of element and software system technology and also the arrival of the time of massive knowledge,deep learning has become this analysis hotspot within the direction of computer science.Convolutional neural network,as one of the typical representative models of deep learning,plays an important role in the field of computer vision.Its features of local perceptual field,weight sharing and down-sampling can realize end-to-end training and testing,thus replacing traditional machine learning algorithms,and has achieved remarkable results in the field of image processing.Image recognition task as a key and representative research direction in the field of computer vision,in the current society dealing with millions of image data,the traditional manual feature extraction methods can no longer meet the basic needs,so the use of convolutional neural network technology can greatly improve the efficiency of image recognition.However,the current convolutional neural network models have the problems of difficult design,low recognition accuracy and huge consumption of computational resources,which are not conducive to the extension to practical applications and services,so there is a need to design more lightweight network models,aiming to greatly reduce the parameter size of the network while guaranteeing a considerable accuracy rate.In this paper,we start from the theory of convolutional neural networks,targeting the current technical problems,and research around the direction of structural design and network optimization of convolutional neural networks to design more efficient lightweight convolutional neural networks,and the main research contents and innovative work of this paper are as follows:(1)Aiming at the problem that the convolutional kernels of traditional convolutional neural networks are too single and do not have diversity,as well as the complex network structure and parameter redundancy,we design a lightweight feature fusion convolutional neural network MS-FNet.The Fusion module of this network architecture uses a multiplex structure to increase the width of the convolutional neural network,and the input feature maps are processed by convolutional kernels of different sizes to improve the network's ability to extract different features in the same layer.It also removes redundant features after each convolution using methods such as batch normalization and Re LU activation function.Using convolutional layers instead of the traditional fully connected layers at the end of the network not only speeds up the training of the model,but also alleviates the problem of overfitting due to too many parameters.Experimental results show that MS-FNet achieves a lower error rate while greatly reducing the number of parameters in the network and has a stronger learning capability.(2)Aiming at the problem that increasing the number of layers of convolutional neural networks leads to a decrease in accuracy,we design a lightweight convolutional neural network incorporating multilevel residual connections MRC-FNet.The multilevel residual connectivity(MRC)module of this network consists of improved residual learning units.Batch normalization,Re LU activation function,and Dropout are used in this residual learning unit to effectively suppress overfitting.Connecting adjacent residual learning units strengthens the information exchange between the front and back layers,realizes the full utilization of features,and ensures that the network can extract richer features.It further alleviates the problem of gradient disappearance and network performance degradation due to the increasing depth of network layers,making the network easier to train.Experimental results show that MRC-FNet achieves a lower error rate while the number of parameters is greatly reduced and the network model has stronger generalization ability.(3)Finally,we discuss the construction,training and testing process of lightweight convolutional neural network models,and deploy the trained network models to practical flower image recognition applications,it has certain application value.Chapter 3 and Chapter4 of this paper detail the lightweight feature fusion convolutional neural network and fused multilevel residual connected convolutional neural network in terms of width and depth improvements,respectively,and achieve good results in public datasets.In order to verify the practical applicability of the proposed algorithm,this chapter utilizes the homemade floral dataset and combines the improvements proposed in Chapter 3 and Chapter 4 to train the model,detailing the data processing process,the model design and training,and the visualization page presentation.The experimental results show that the lightweight convolutional neural network model proposed in this paper has a high recognition rate and is capable of performing basic recognition tasks.