Classification model based on small sample image data -- simple version

How to perform the task of image classification with none few samples in each category requires three steps:

1. Crawling data,
2. Training coarse-grained classification model
3. Training refined classification model

Data crawling using beautifulsop processing

Code data processing adoptstorchvisionprocessing

This model adopts the trained version of vgg16 model of pytorch

Crawling image data to be classified

Core code：spyder/ImagesSpyder.py

Fill the category information to be crawled into Spyder/name.TXT

After running, input the number of pictures to be crawled for each category

Data handling

Select the crawled pictures and put them in the directory of ./data/train_old/, remember to store them according to the principle of one directory for each category

Generally, it can be:

./data/train_old/0_person/
./data/train_old/1_flower/
...

Put the precious small amount of sample data you have into the directory ./data/train/, which is the final classification target

Model training

The core file isrecognition/models.py

What needs to be revised:

1. Modify n_classes in the main function to the number of your categories
2. Model address model_base_path
3. Your GPU info : os.environ[CUDA_VISIBLE_DEVICES]=0,1``

As the topic said, you need to use different training sets to train the model twice. For the first time, you need to use the data in train old, and for the second time, you need to use the data in train

Remember to switch datasets when using. By default, you can use datasets in train for training. You can modify them freely

Test Result

The following is the accuracy test results. Of course, we should not only focus on the accuracy, but also pay attention to the sample situation. It is better to use Gan to generate small samples to increase the sample size...

Base Model：

Update Model：

Download&Contact

Basic model download:：VGG16

In case of failure, contact: emsunfc@163.com