Research And Application On Image Recognition Of Chinese Food Based On Deep Learning

Chinese dishes image recognition has been a welcome area with the tech of computer vision flourishing,and has extensive and important application in many fields,such as mobile devices,health assessment,etc.With the rapid development and maturity of deep learning technology,Chinese food image recognition has also made breakthroughs.However,due to the complexity of Chinese food images and background noise,it’s important to know the way to expeditously and precisely handle feature extraction from food images,efficiently detect multi-target Chinese dishes and calorie prediction of food based on images,there are still many huge challenges.To solve the problems,we did the following work:For the single-target Chinese food image recognition,a feature extraction and learning algorithm based on attention functional network is proposed,which can advance the capacity to extract distinguishable semantic features of images and learn features of pivotal regions,deminish image noise,and enhance features significance.Comparing experiments with the classic network structure,this algorithm expeditously ehances the recognition accuracy,and the accuracy rate in the supplemented VIREO Food-172 data set is 87.6%.For the multi-target Chinese dishes image recognition,based on the Mask R-CNN,an algorithm based on depthwise and pointwise convolution mechanism is proposed,and its backbone network is modified and optimized.Finally,a comparative experiment is conducted on the COCO data set and the Chinese dishes image target detection dataset.The evaluation of comparing model size and training time consumption shows that the modified Mask R-CNN can significantly diminish computing resources,and to a certain extent,it can avoid performance degradation arising from the go-up in network complexity,improve detection precision and can ensure accuracy.For the food calorie prediction task,a food calorie prediction algorithm based on energy distribution is proposed,and the modified Mask R-CNN is used to detect food and alignment substances and classify them,which expeditously advances the detection accuracy and diminishs the error in food calorie assessment.After related experiments on the ECUSTED dataset,it is concluded that when the IOU threshold is 0.75,the detection classification of all categories achieves the best effect,and compared with the calorie prediction method mentioned above,the error rate of the algorithm proposed in this paper is diminishing over 10%,it is verified that the algorithm has superior performance in calorie budget.