Text Detection Of Food Labels Based On Semantic Segmentation

The text of food labels as food-related information provides the important foundation for consumers to buy food,it also provides the possibility for food supervision and sampling administrations to carry out data mining and discover potential food safety problems.Text detection of food labels can help to decrease the heavy workload of manual inputting food label texts and advance the efficiency of food safety data analysis and mining for food supervision and sampling administrations.Therefore,as a necessary step of recognition of food labels,the detection of food labels is of great significance.Firstly,research status of text detection based on deep learning is summarized and classified in this paper.The model structure,technical route,advantages and disadvantages of the main model methods are expounded.Secondly,to achieve the text detection of food labels,the dataset of food packaging is constructed,and a semantic segmentation based distance field model is proposed,in which two tasks are included:pixel classification and distance field regression,the pixel classification task is used to segment the text and background regions,and the distance field regression task is used to predict the normalized distance from the pixel located in the text region to the boundary of text region.In order to effectively use the correlation of two tasks to improve the detection performance of the model,an attention module is added into the distance field regression to optimize the model structure,in addition,in order to prompt the experiment accuracy,the loss function is improved to solve the problem that the loss value of the distance field regression is too small.The results of ablation experiment show that the accuracy of the proposed model is increased by 4.39%and 3.80%respectively according to the improvement of attention module and loss function.The comparative experiments of different model methods show that our model has good performance in detecting the text of food labels,and the recall,accuracy and F-measure are 87.61%,76.50%and 81.68%respectively.Finally,the proposed model is applied to the service of text detection and recognition of food labels.The detection results are input into the text recognition model to obtain the text of labels.To reduce the recognition error,a rule-based text correction is implemented.Finally,the whole module of detection and recognition encapsulated into an API to provide the recognition service for clients.