Research On Dishes Recognition Algorithm Based On Few Sample Learning

Since deep learning has become a hot research direction,various research methods based on deep learning have also emerged in an endless stream.Since machine learning was proposed in the 17 th century,many methods have been developed.However,the emergence of deep learning has changed the traditional solutions and made machine learning more malleable.Dishes recognition as an important application in the field of computer vision in real scenes has a broad research content.Dishes identification can not only be used for automatic pricing in the catering industry,increase the settlement window,reduce human resource loss,and reduce customer waiting time through the rapid calculation of the machine,but also be used for "scanning" in applications such as We Chat to identify each Dishes.In the dish recognition method,traditional image processing methods and machine learning methods have many shortcomings.Images taken in actual scenes will be interfered by external factors such as environmental light intensity,noise interference,ambient light,etc.,resulting in poor quality of the captured images,which will affect the final detection results.With the development of deep learning,convolutional neural networks have achieved great success in image recognition,target detection,semantic segmentation and other fields.Dishes recognition has also been researched around convolutional neural networks.Small sample learning has received a lot of attention since its development.The main research content of small-sample learning is to obtain good test results on a small number of data samples,so that it can better adapt to today's rapidly changing society.When applied to actual dish recognition,both recognition accuracy and detection speed need to be taken into consideration.Few-shot learning has received a lot of attention since its development.Deep learning is based on a large number of sample data sets,which limits the practical use of deep learning and its application in rapidly changing environments.Therefore,few-shot learning came into being.It can get good detection results on a small number of sample data sets,and it is more adaptable to today's rapidly changing society.Considering the diversity of dishes,the large number of shootings of each type of dishes not only increases the cost,but also makes it difficult to update the dishes.Both the collection of sample data sets and the training of models require a lot of manpower and time.At the same time,the application of few-shot learning to actual dish recognition needs to balance recognition accuracy and detection speed.This article is dedicated to solving the above problems and studying the identification of small-sample dishes in actual restaurant application billing scenarios.The main work includes the following aspects:(1)Construct and open source the small-sample learning dish data set Food-270,which is used for dish classification research.The Food-270 data set is collected from real restaurants,including 270 types of dishes,divided into a support set and a verification set.Each type of dish in the support set includes 20 sample images,and each type of dish in the validation set includes 20-30 images for research.The generalization ability of dish recognition in the case of a few samples in the actual restaurant.(2)Explore the influencing factors of image retrieval and clustering on dish recognition with few samples.The image retrieval speed is fast,but the detection accuracy is not high when there are few samples.Use methods such as data enhancement to expand the image retrieval database;change the distance function to make the distance function suitable for dishes;re-search the retrieval results to improve the final retrieval accuracy.Different clustering methods are suitable for different research occasions.According to the special situation of dishes,the influence of different clustering methods on the detection results is studied.(3)Propose a small-sample dish recognition algorithm based on class difference and contrast learning.Contrastive learning,as a method of self-supervised mode,has good accuracy in image recognition.For the first time,this thesis applies ratio learning to the small sample recognition of dishes,and for the unequal number of samples for each dish in actual use,a small sample dish recognition algorithm based on class difference and contrast learning is proposed to improve the contrast.Learn the detection accuracy in the case of few samples.(4)Apply Transformer to the identification of dishes with few samples.Transformer was first proposed in the field of natural language processing.It is formed by connecting encoding components and decoding components.The network itself has a strong self-attention structure.This article discusses the use of Transformer in image recognition,and applies it to the recognition of dishes with few samples for the first time,and obtains better detection accuracy.