Research On Structure Optimization Algorithm Based On Convolutional Neural Network Model

Convolutional Neural Network(CNN)has been widely used in the field of computer vision for their excellent feature extraction ability,as it is supported by today’s large-scale computing power and training data,but considering that it is difficult to deploy larger models in practical applications such as mobile devices due to the limited storage and computing power,thus the structural optimization algorithms based on convolutional neural network models have become a popular research topic.The main solutions are various types of model compression strategies,lightweight model structure design and attention mechanism to optimize the convolutional neural network model.In this paper,we optimize the performance of the convolutional neural network model in terms of memory consumption,storage,computation and model accuracy from two aspects: the convolutional structure of the lightweight model and lightweight attention mechanism.The main work and contributions of this paper are as follows:(1)A lightweight convolution module based on feature redundancy,namely spatial and channel reconstruction convolution(SCConv),is proposed.The SCConv module is divided into two parts,namely,spatial reconstruction unit(SRU)and channel reconstruction unit(CRU),to reduce the spatial and channel redundancy of the feature map and improve the capability of feature representation,which reduces the number of parameters and computational cost required for feature extraction and makes the model more compact.It also improves the accuracy of the model.Meanwhile,the SCConv module is a plug-and-play module that can be embedded into CNNs replacing the standard convolution,and it can achieve better results in different model architectures.(2)A weight attention(WA)module based on feature mapping is proposed.The WA module is a lightweight and efficient plug-and-play attention module that almost adds no additional parameter burden and computational overhead to the network,and can be embedded into any convolutional neural network models as a beneficial module to improve the accuracy and speed of the model.The proposed method is validated by experiments on three image classification datasets:Image Net-1K,CIFAR-10 and CIFAR-100,and the experimental results show that the proposed method can optimize the performance of the computational storage and classification accuracy of the convolutional neural network model in terms of both lightweight convolutional structure design and lightweight attention mechanism,and it can achieve better results in comparison with related state-of-the-art methods.The results show better performance in comparison with the latest methods.The experimental results on the PASCAL VOC and MS COCO target detection datasets demonstrate that both proposed methods have good migration and generalization performance and can be extended to a wider range of applications.