Research And Implementation Of Deep Learning Dish Image Recognition System For Catering

With the rapid development of artificial intelligence,artificial intelligence applications have made great progress and entered people’s daily life,and diet is an essential part of people’s daily life.The identification and classification of dishes are currently attracting much attention.A field of research.Secondary developments around the identification and classification of dishes,such as food health management,have brought great convenience to people’s lives.The identification and classification of dishes can be regarded as a target detection task with industry characteristics,but the manual features and classification methods in traditional target detection algorithms can no longer meet the requirements of detection accuracy and detection speed.Therefore,deep learning-based target detection is required.Methods The field of dish identification and classification was introduced.Combined with the actual application scenario of the project,compared with traditional object detection tasks,the classification and recognition of dish pictures has more difficult points such as users taking pictures that are greatly affected by light,strong background noise,diverse picture resolution and target scales,and large gaps between dish picture categories.In view of the above problems,this article has done the following work based on previous research and combined with practical applications:(1)For image illumination and background noise,image pre-processing is performed before image recognition and classification.In this paper,CLAHE limited contrast histogram equalization algorithm is used to limit the effect of image illumination on image characteristics.Wavelet transform filtering algorithm with improved threshold function is used.Reduce image noise,reduce the impact of image lighting and complex backgrounds on dish recognition and classification through image preprocessing,and improve the quality of image to be processed and the efficiency of model detection.(2)An improved SD feature extraction network based on Dark Net is proposed.S3 Pool random space sampling pooling is used to replace the pooling layer in the Dark Net feature extraction network.Random space sampling pooling eliminates the traditional pooling layer for image features.Loss,while expanding the model training parameters to improve the generalization ability and detection accuracy of the model.(3)Based on the YOLOv3 model,an improved MSSD-YOLOv3 multi-scale enhanced network model is proposed.Based on the three-layer detection scale of the original YOLOv3 model,this paper builds a multi-scale enhanced network model by expanding small and medium-scale detection to achieve It predicts pictures at 5 scales,improves the detection accuracy of small target dishes in actual application scenarios,and uses k-means clustering algorithm to re-cluster the dish picture data set to get suitable for dish picture detection tasks.The detection prior frame improves the model detection accuracy and speeds up the model detection speed.At the same time,the proposed model model is optimized by using the merged convolution layer and BN layer technology and batch re-normalization technology,reducing the model calculation parameters and increasing the detection speed.The training and detection effect of the model on small samples and dependent independent distribution samples.(4)Designed multiple sets of control experiments for the models before and after the improvement.Based on the analysis of the experimental results,the detection accuracy and speed of the dish image recognition and classification model proposed in this paper are improved compared to the traditional dish detection algorithm and YOLOv3 model.Analyze and compare the experimental results and parameter selection.(5)Finally,the proposed dish identification and classification model was combined with the Gin microservices distributed framework,React front-end framework,OSS object storage,Redis and My SQL database to develop a complete dish image classification and recognition system and open API interface.In 4C8 G The standard server provides more than 1200 QPS concurrent access capabilities,provides technical support for other users and the secondary development of dish classification technology,and achieves good results.