Research On Feature Fusion Algorithm Based On Lightweight Network

Convolutional neural network(CNN)is widely used in the field of computer vision and has achieved good results.With the increasingly close connection between computer vision and pattern recognition and life,in some real application scenarios such as mobile terminal or embedded devices,large and complex models are usually difficult to be applied.Therefore,it is very important to study a small and efficient CNN model.The lightweight network is suitable for real-time classification or detection due to its small number of parameters.Although the current lightweight network has made great progress,there are still the following problems: 1.The main structure of most networks is relatively simple,and feature information in different feature layers is not fully integrated.2.The system performance is generally low,and there is no good balance between speed and accuracy.In view of the above problems,this thesis conducts in-depth research on the feature fusion algorithm in the lightweight network,and makes a series of improvements to the network structure of the lightweight network.It mainly includes: 1.A feature fusion network structure based on Mobile Net light network is proposed,which effectively improves the accuracy of image classification.2.The feature fusion module Middle Fusion Block is proposed,which adds the output results of each Block in the backbone network layer by layer after scaling,achieving a good fusion effect.3.The attention module is improved,and the effect of target recognition is improved by adding parallel spatial attention module in addition to channel attention.4.The module FFM for upper-lower feature fusion is proposed to realize the effect of feature integration relearning and improve the accuracy of target recognition with a small amount of calculation.We have verified ZNet network structure,attention module DAM and feature fusion module FFM on CIFAR-10,CIFAR-100,SVHN and other data sets in large numbers.Our experiment shows that compared with other lightweight classification networks,our network structure has been improved in terms of accuracy and speed.At the same time,the effectiveness of the network in case segmentation proves that the feature extraction ability and speed of the network can be applied to real application scenarios.