Research On Food Logo Detection Method Based On Deep Learning

With the continuous development of deep learning technology,convolutional neural network has been widely used in many scientific research fields,and deep learning technology has become one of the effective means to solve problems in the field of computer vision.At present,object detection plays a great role in intelligent video surveillance,robot navigation,aerospace and other fields,and is widely concerned by researchers.With the advent of the era of big data,the amount of data increases exponentially,and the update and iteration of hardware facilities such as graphics processing unit(GPU)drives the rapid development of object detection algorithms.As a branch of object detection,logo detection from images is one of the most distinctive and effective methods to determine the brand.However,logo detection is still a challenging problem due to large differences in image size,geometry,appearance and shooting angle.In reality,there are many kinds of brand logos.The food logo is the most common one in life.Food logo detection has a wide range of applications in real life,such as food recommendation in self-service stores,infringement detection in e-commerce platforms and targeted advertising.Compared with common logos,food logos have more interclass similarity and more complex multi-scale problems.In this thesis,food logo detection is carried out by constructing a food logo dataset and food logo detection algorithm,which can be divided into the following three steps:(1)Data quality is an important basis for deep learning algorithms to play a role.In order to develop advanced food logo detection algorithms,large-scale food logo datasets are needed as support.However,there is no publicly available food logo dataset.To this end,this thesis constructed the Food Logo Det-1500 dataset,which is the largest publicly downloadable food logo dataset with 1,500 categories,approximately 100,000 images,and approximately 150,000 handlabeled food logo objects.This paper introduces the collection,cleaning and annotation process of the dataset in detail,analyzes the size and diversity of the dataset,and compares it with other logo datasets.According to the research,Food Logo Det-1500 is the first of the largest publicly available high-quality dataset for food logo detection.(2)In this thesis,a Multi-scale Feature Decoupling Network(MFDNet)based food logo detection model is proposed to accurately detect food logo.The method decoupled the classification task and the regression task into two branches to solve the multi category detection problem of food logo.Specifically,the Feature Offset Module(FOM)is introduced,which uses deformation learning to obtain the optimal classification offset,and can effectively obtain the most representative classification features during detection.In addition,Balanced Feature Pyramid(BFP)module is also used,which focuses on global information and enhances feature extraction ability.Moreover,the fusion of multi-scale features is further enhanced to enhance detection ability of multi-scale objects.(3)Finally,an extensive experimental evaluation is performed on three datasets,including the proposed Food Logo Det-1500 and two other widely used logo datasets,QMUL-Open Logo and Flickr Logos-32.Experimental results verify the validity of Food Logo Det-1500 food logo dataset and the feasibility of the proposed method.