Research On Detection Algorithm Of Chinese Dishes Based On Deep Learning

With the improvement of computer hardware performance,Artificial Intelligence technology has made great progress,and its application has gradually entered people's daily life.Food is an inseparable part of people's lives.In recent years,research on dish detection technology has gradually become hot.The secondary development on the basis of the dish detection,is applied to the fields of automated catering services,healthy diet management and etc.,which provides great convenience to people's lives.At present,the work of serving food in the restaurants and school canteens is boring,and requires large labor demand and high cost.It has become a solution to replace people with robots.The first step of dish-serving is to identify and locate the dishes,so the algorithm research for the detection of Chinese dish is of great significance.Based on the One-stage object detection method of deep learning and combined with the application of Chinese dish detection,this thesis optimizes data and image processing and network structure design.The main contents are as follows:1.Establishment of Chinese dish data set.The current public dish data set is an image classification data set,and the image only contains a single dish,which is not suitable for object detection.This thesis constructs a Chinese dish data set of containing 37 categories,and each image contain many different dishes.Then compared with the public dish data set and object detection data set,it verified the rationality and superiority of the Chinese dish data set in this article.2.Preprocessing of dish image.In view of the background noise and uneven illumination in the dish image,the dish image should be preprocessed.In this thesis,Non-Local Means technology is used to reduce the noise of the dish image,and the Local Homomorphic Filtering algorithm is used to handle uneven illumination,while enhancing the detailed features of dishes.Image preprocessing improves the accuracy by optimizing the image quality.3.Design of dish detection network.Aiming at the problem of image feature loss caused by pooling or 2-step convolution,the image slice method with interval sampling and stacking is used to better retain the detailed information of the image while down-sampling.The residual unit structure with channel attention mechanism is proposed,and the design of the feature extraction network is completed by combining the Cross-Stage Hierarchical merging structure,which strengthens the feature extraction ability and reduces the amount of parameter calculation.Add fusion patht to make good use of detailed and semantic features to complete the detection of dishes.4.Processing of prediction boxes.Aiming at the local repetitive detection problem in the detection of Chinese dishes,a new prediction box processing strategy is proposed.A new prediction box distance calculation method is used to suppress the prediction boxes that appear to contain the relationship.This thesis conducts experiments on the above work,and the results validate the effectiveness and advantages of the method proposed.The detection mean average precision in this thesis reaches95.1%,and the detection speed reaches 24 fps,which achieves the desired effect.