Specific Logo Detection Using Few Samples

In recent years,with the rapid development of computer vision and the continuous deepening of network technology,object detection has also developed rapidly as an important research issue in the field of computer vision.Logo detection as the sub-problem of object detection has also received full attention from researchers.The problem of logo detection has been widely developed and applied in the fields of commercial advertising,copyright infringement detection,and intelligent transportation systems.Logo has a specific and highly representative form,with multiple transformations under different scenes and attachments,posing a serious challenge for this task.With the advent of the era of big data and the in-depth study of deep learning techniques,the deep learning-based convolutional neural network model has achieved remarkable results in the field of computer vision applications.However,due to its strong dependence on the training data,it brings challenges to the visual field tasks of some small sample problems.Therefore,it is of certain research and application value to study the logo detection problem under small samples.This thesis introduces the traditional logo detection method first,and then introduces the current mainstream deep learning-based object detection framework.It proposes a basic framework for few-example detection and explores the detection algorithm with more model expressiveness and then applies to the task for a specific logo dataset.The main works are as follows:(1)According to the characteristics of the small sample logo detection task,a three-stage logo detection framework suitable for few-example logo detection is constructed based on the idea of migration learning.The first stage uses the basic logo detector to pre-train on ImageNet to initialize the parameters;the second stage trains the step-by-step refined model with a large number of synthetic images;the third stage trains with a small number of real logo examples for the refined model.(2)From the data level,we improve the few-example logo detection problem by constructing context-based synthetic samples.According to the environment attribute of the logo type,select the appropriate identification template,perform various image processing transformations,insert the random position into the background image,and finally perform the Poisson fusion method to make it more realistic with the background fusion.(3)From the model level,based on the characteristics of the logo,the detection algorithm with more model expressive power is explored,so that it can make greater use of its advantages under few examples.First,we use Faster R-CNN for logo detection.This method uses the RPN and Fast R-CNN to share the convolutional features;and then insert deformable convolutional network to further enhance the ability to deformation and scale of objects.