Research And Application Of Lightweight Object Detection Network Based On Feature Enhancement

In recent years,detecting food with computer and using artificial intelligence has replaced traditional methods such as manual persuasion and identification,becoming a new way for canteens in universities and businesses to promote the clean plate activities.This emerging approach,combined with various incentive activities such as checking in with clean plates and exchanging gifts,has received a positive response from university students and others.However,simple models are difficult to generalize to diverse plate types,while models with large parameter sizes can achieve better accuracy but face deployment difficulties due to their slow inference speed and large memory usage,resulting in a poor user experience.To address the issue of slow inference speeds and accuracy drop after model light weighting,this paper proposes a lightweight object detection model based feature enhancement.First,a tiny version is used to reduce the number of convolutional layers,eliminate multiple fusion layers and detector,and decrease the computational cost to improve inference efficiency.Then a coordinate attention mechanism module is introduced to improve the network’s feature extraction ability.This is achieved by designing and adding an output from the last convolutional layer of the backbone network,and an output from the up-sampling process in the feature fusion layer,so that the network focuses more on the relevant parts.The Res Block module is replaced by a feature extraction enhancement module designed in this paper,which includes a basic convolutional layer for feature extraction,soft-pooling to improve the loss of information during image down-sampling caused by ordinary pooling,and a lightweight activation function to optimize smoothness while reducing computational costs.In addition,this paper proposes a practical solution for model deployment,where the trained model is deployed as a service and published on a mini-program.The model can also achieve inference in seconds on a single-core CPU,providing a reasonable solution for similar requirements in practical applications.Based on the above design,this paper proposes a lightweight object detection model.A lot experiments are designed.On our dishes dataset,the experimental results show that the proposed model achieves an m AP of 90.51% and a frame per second(FPS)rate of 93.12,which is 4.47 times faster than Faster R-CNN with a slight difference in accuracy.By introducing the Coordinate Attention Mechanism,the m AP value is improved from 87.26% to 88.33%.Moreover,replacing two different Feature Enhancement Modules can result in 2.08% and 3.02% performance improvements,respectively.The experimental results show that the proposed method achieves a balance between FPS and m AP,outperforming some models in both aspects.It is also more suitable for practical deployment and application and has greater advantages in meeting design standards.