| Deep convolutional neural networks have shown outstanding performance in image recognition.However,as the network size continues to increase,the required computing and storage resources sharply rise,making it difficult to perform efficient image recognition in resource-limited environments such as mobile devices.The emergence of lightweight convolutional neural networks effectively solves the problem of highprecision complex models relying too much on high-performance computers,but there is a slight loss in recognition accuracy.This thesis conducts research on reducing network size and ensuring recognition accuracy,and proposes an optimized lightweight network model structure and loss function.Experimental results on multiple datasets demonstrate the effectiveness of the optimization.The main research work of the thesis is as follows:1.Six classic lightweight neural networks,Mobile Net V2,Shuffle Net V2,Squeeze Net,Xception,Efficient Net,and Ghost Net,were studied and compared in terms of recognition accuracy,precision,recall,model parameters,computational complexity,processing speed,etc.Based on the experimental results,Ghost Net with good comprehensive performance was selected as the object for subsequent research and optimization.2.Based on the Ghost Net model,an optimized model L-Ghost Net is proposed.Learning group convolution and an improved channel attention mechanism are integrated into Ghost Net,reducing model complexity and the number of parameters.At the same time,the pruning ratio is added to learning group convolution to control the end time of pruning in the entire process.The improved channel attention mechanism replaces the convolutional layer with a fully connected layer,making the connection between the two dimensions tighter and increasing the model’s flexibility.Experiments on multiple datasets in various fields,such as Grape leaves recognition,Gesture recognition,face recognition,and Rice recognition,show that L-Ghost Net has slightly improved accuracy,reduced computational complexity by over 44%,reduced the number of parameters by over 33%,and improved FPS by 33% compared to Ghost Net.Compared with other commonly used lightweight network models such as Mobile Nets and Shuffle Nets,LGhost Net has the lowest FLOPs,the highest accuracy,and fewer parameters at the same computational complexity,with good comprehensive performance.3.Based on a variable margin loss function,an Mp Loss loss function model is proposed.The model adds hyperparameters m and parameter p,where hyperparameter m is used to control the threshold size between different classes,and parameter p records the model’s training progress.Using L-Ghost Net as the experimental model,experiments on multiple datasets in various fields,such as Grape leaves recognition,Gesture recognition,face recognition,and Rice recognition,show that Mp Loss has slightly improved accuracy,precision,recall,F1-macro average,etc.,compared to other loss functions such as Center Loss,L-Softmax,A-Softmax,Cos Face,Arc Face,and Mag Face,with good comprehensive performance. |
