Shulou(Shulou.com)06/02 Report--

2019-10-11 12:38:37

The column "Image Classification" is officially over! We explain from the data set, from the most basic multi-category image classification to fine-grained image classification, multi-label image classification, and then to the more difficult unsupervised image classification. then we describe the various problems faced in image classification and comprehensively summarize the relevant content in the field of image classification. At this point, we will make a big summary of the content of the whole image classification column!

Author | Guo Bingyangyan has three

Editor | there are three words.

1 Symphony of datasets and classic networks

As a data-driven tool, deep learning needs a large number of high-quality data sets to carry out the corresponding work. We summarize in detail the data in various fields of image classification at this stage, and point out its applicable classification tasks, which provide a compass for readers to carry out the corresponding work.

At the same time, we summarize the classical image classification network and introduce in detail the major breakthroughs and typical design ideas in the history of image classification.

The classification column of "Image Classification" is officially online! For new entry into CV and AI, you need a compass

Image classification starts with data sets and classical networks.

2 multi-category image classification

Multi-category image classification is the most basic kind of image classification, that is, the image with a single category of objects is classified. Multi-category image classification has been integrated into all aspects of life and has been successfully applied to many fields such as face recognition.

To build a complete image classification task, we need to select an open source framework, build and read data sets, build a network framework, train and debug parameters, and test accuracy. This paper uses pytorch open source framework as a tool, white blood cells as a dataset to complete a multi-category image classification task.

"Image Classification" Multi-category Image Classification based on Pytorch

3 multi-label image classification

Pictures in real life often contain multiple categories of objects, which is more in line with people's cognitive habits than multi-category image classification.

Multi-tag image classification can tell us whether the image contains these contents at the same time, which can also better solve the problems in real life.

In many solutions, the idea of multi-category image classification can be quickly transferred to multi-label image classification by tag transformation. In the actual combat part, with the help of this idea, we use ResNet 50 as the basic framework to achieve multi-label image classification based on pascal voc data sets.

A Summary of Multi-label Image Classification in "Technology Review"

4 fine-grained image classification

Fine-grained image classification is on the basis of distinguishing the basic categories, making a more fine subclassification, such as distinguishing the species of birds, the style of cars, the breeds of dogs, and so on. At present, there are a wide range of business needs and application scenarios in industry and real life.

Compared with coarse-grained images, fine-grained images have more similar appearance and characteristics, coupled with the influence of posture, viewing angle, illumination, occlusion, background interference and so on, which lead to the phenomenon of large differences between classes and small differences within classes, which makes the classification more difficult.

In the actual combat part, we complete a basic fine-grained classification task with CUB data sets through the classical bilinear network structure.

"Image classification" what is the fine-grained image classification, what are the methods, and how is it developing?

"Image Classification" Fine-grained Image Classification based on Pytorch

5 unsupervised image classification

At this stage, the task of image classification is largely achieved by supervised learning, that is, each sample has its corresponding label, through the depth neural network to constantly learn the corresponding features of each label, and finally achieve classification. In this case, the capacity of the data set and the quality of the label often play a decisive role in the performance of the model.

High-quality data sets will naturally bring difficulties in labeling and cost a lot of human and financial resources. In this case, how to reduce the label information and ensure the classification accuracy has also become a further research goal, which is also the development direction of unsupervised image classification.

"Image Classification" briefly describes the development status of unsupervised image classification

6 categories unbalanced

Category imbalance refers to the large difference in the number of training samples of different categories in the classification task, which is usually caused by the difficulty of sample collection or the lack of sample examples, which often appears in disease category diagnosis, fraud type discrimination and other tasks.

This article will summarize the relevant solutions involved at present, which are divided into three aspects: data level, algorithm level, data and algorithm mixed level, and enumerate representative schemes for readers' reference.

"Image Classification" about the imbalance of categories in image classification

7 label noise

In the data set production process, due to subjective and objective reasons, it will lead to the emergence of label noise, and its existence will have a certain impact on the performance of the final classification model. Therefore, in the process of practical application, it is always necessary to clean the existing data set to avoid the interference of label noise.

In this paper, the solution of label noise is introduced, and the actual combat is carried out by iterative training. At present, there is not a completely efficient and proper way to solve this problem, and it is still in a stage of research and development.

What is the impact of "image classification" label noise on classification performance?

8 fight against attack

Most of the image classification networks based on deep learning are trained under elaborate data sets, and the corresponding deployment is completed. For images outside the data set or slightly modified images, the recognition ability of the network is often affected to a certain extent.

Today, when more and more attention is paid to security information, how to improve the robustness of the network model to noise, interference and other antagonistic samples is the further research goal of image classification.

What is the counterattack in "image classification" image classification?

9 visualization

Think the deep learning model is like a black box? Can't understand how it works? Then visualization is done, here we give you a summary of the visualization of image classification models and practical cases.

Visualization of Image Classification Network in practical Operation

"Technology Review" and "see through" Neural Network

10 Overview of image classification

In addition, we have summarized a summary of image classification to help you understand the principles of image classification at different levels, which is also a summary article before the start of this column.

"Technology Review" do you really know anything about image classification?

11 image classification model architecture

Last year, we also summarized the interpretation of the architecture of 12 mainstream deep learning models, most of which are image classification models for beginners to facilitate systematic learning. If you want to know more about the strange model architecture, you can apply to join the three AI knowledge Planet, hundreds of issues waiting for you to read.

"the end" summarizes the architecture design ideas of the 12 mainstream CNN models.

12 mainstream open source framework classification projects

In order to make it easier for everyone to use, there are 12 deep learning open source framework image classification projects of three AI, from data preparation, model definition, visualization, training and testing, graphics and text tutorials and complete code, what are you waiting for?

"end" 12 deep learning open source framework quick start projects for beginners

Summary

After several months, under the guidance of my own efforts and three elder brothers, the image classification column has finally come to an end. I hope you can get something in the column and get a little promotion. At the same time, thank you very much for your comments and guidance, which have benefited me a lot.

This period of time for myself is a stage of self-improvement and continuous breakthrough. Not only consolidated their basic knowledge, but also gained a lot of experience. These experiences have made me understand that if we do things in a down-to-earth manner and keep working towards our goals, there will be gains.

Finally, I would like to give you the passage at the beginning of the column:

Where there is a will, there is a way. Two hundred Qin passes belong to Chu in the end.

Painstaking people, days live up to, three thousand more A can swallow Wu.

Https://www.toutiao.com/i6746400270978122253/

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.