Image Recognition And Segmentation Algorithm Based On Mask R-CNN

With the development of food culture information,a variety of food image sharing websites have developed rapidly,such as Bean Fruit Food Web,World Food Web,Public Comment Web,etc.Lots of similar websites appear in people’s field of vision.How to get practical information from a large number of dishes is of great significance in practice.According to their own need,effective segmentation and recognition of dishes can help people to efficiently search,analyze and summarize in a great deal of online food images.For example,people with diabetes can avoid all the foods with high sugar content;Hui people can block all the ingredients of pork raw materials;doctors can also understand the dietary structure of patients before the information,and give reasonable nutritional advice.In short,the automatic segmentation and recognition of food images not only has practical significance for people’s diet health and quality of life but also systematically promotes Chinese food culture through the network.This paper mainly uses the target segmentation model Mask R-CNN in deep learning to segment and identify dishes.First,the crawler technology is used to obtain the image data of the dish,and each strategy and rule of the web crawler is compared to select the most suitable crawling technology.Then,data cleaning and filtering of the acquired data is used to ensure the quality of the data.Thereafter,images are labeled,the processed data is divided into training set and test set,and a dish data set is constructed.Then deep separation convolution improves the Mask R-CNN model,reducing the amount of model calculation and storage space.After evaluation of the COCO dataset,it found that replacing the standard convolution of the Mask R-CNN model feature extraction layer with the deep separation convolution can achieve the desired effect.Finally,the constructed dish training set is used to train our model and improve Mask R-CNN.The testing set is then used for verification.According to the comparative analysis of models,it concludes that the improved model provides much faster and powerful calculating speed,and reduces calculation as well as computing resource,at the cost of a slightly reduction of accuracy.The experimental results show that the algorithm can significantly reduce the resource consumption of the Mask R-CNN model.Besides,the efficiency of segmentation and recognition also can be improved.