The latest 2022 review of "small sample deep learning image recognition"

Image recognition is the core problem in the field of image research , Solve the problem of image recognition, face recognition 、 Autopilot 、 Research on robots and other fields is of great significance . At present, the widely used machine learning method based on deep neural network , It has been classified in birds 、 Face recognition 、 Daily object classification and other image recognition data sets have reached a level higher than human beings , At the same time, more and more industrial applications begin to consider the method based on deep neural network , To complete a series of image recognition services . However, deep learning methods rely heavily on large-scale annotation data , This defect greatly limits the application of deep learning method in practical image recognition tasks .  To address this issue , More and more researchers begin to study how to train the recognition model based on a small number of image recognition annotation samples . In order to better understand the problem of image recognition based on a small number of labeled samples , Several mainstream annotation learning methods in the field of image recognition are widely discussed ,  Including methods based on data enhancement 、 Methods based on transfer learning and methods based on meta learning , By discussing the processes and core ideas of different algorithms , We can clearly see the advantages and disadvantages of the existing methods in solving the problem of image recognition with a small number of annotations . Finally, aiming at the limitations of existing methods , The future research direction of small sample image recognition is pointed out .

Address ：

http://www.jos.org.cn/jos/article/abstract/6342?st=article_issue

Today's machine learning methods ,  In particular, the machine learning method based on deep neural network has been used in face recognition [1]、 Autopilot [2]、 robot [3] And other image recognition related fields have made great achievements ,  Some have even exceeded the current human recognition level .  However, great achievements have been made in deep learning At the same time ,  People find it difficult to apply it to practical problems .  The first is the problem of labeling data ,  Current deep learning methods require a large number of standards Note data to train [4] ,  However, in practical application, data acquisition is often difficult ,  There is also the issue of personal privacy ,  Like face data ,  Also have The problem object itself is very few problems ,  For example, the problem of identifying rare and protected animals ,  besides ,  Data annotation often requires a lot of manpower and material resources ,  This hinders the application of deep learning technology in the field of image recognition .  The second is the problem of computational power ,  The deep learning method can improve the performance of the algorithm at the same time ,  Often accompanied by huge network operations ,  This makes it difficult for deep learning methods to be deployed on devices with limited computing resources ,  Therefore, in some cases, the computational power is limited Application scenarios of ,  Like autonomous driving 、 robot 、 Road monitoring and other issues ,  At present, most of the image recognition tasks still use some low intelligence 、 Technology with low computational power consumption ,  This also seriously hinders the development of intelligent image recognition technology . 

By contrast ,  Human recognition is relatively light ,  That is, you don't need to collect a lot of data to learn ,  It doesn't take a long time to think or Calculation [5] .  For example, parents teach newborn babies to read ,  Distinguish animals ,  Simply paste one or two corresponding calligraphy and paintings at home ,  Children will soon recognize Read the content above .  How to retain the powerful knowledge representation ability of the current deep learning methods at the same time ,  So that it can quickly learn from a small number of samples Useful knowledge ,  This problem of image recognition based on small samples has gradually attracted people's attention .

This article will discuss in the following order ,  First, in the first place. 1 This section introduces the problem description of small sample image recognition ,  Then it will be on the 2 This section introduces the base Small sample learning algorithm for data enhancement ,  In the 3 This part introduces the algorithm based on transfer learning ,  In the 4 This section introduces the algorithm based on meta learning ,  It will be in the 5 This section introduces the widely used evaluation indexes of small sample image recognition ,  And compare the performance of the above algorithm on the benchmark of the problem ,  In the end The first 6 Part refers to the shortcomings of existing algorithms and the future development direction .