• Please make sure to install tensorflow, keras, cv2, numpy and matploblib.
• Please change those directories in the head of image_classification in it into your own path.

This program can automatic build and train a 3/5 lays convolutional neural network to classify any kind and any number of objects you want to recognize using tensorflow and keras based on datasets you provide.

You need to prepare 3 datasets: Train, Test and Validation. Each of them should contain several folders and each folder shoud contain a kind of object you want to classify and the folder will be use as label for images in it.

You can run it directly after preparing those datasets, and it will automatically build and train a 3 lays convolutional neural network and a 5 lays convolutional neural network and solve the best model as best_model3/5.hdf5. After finishing, you could either choose some test images and show the prediction results using test_plot or plot the Training and validation loss and Training and validation accuracy's trend into line chart using evaluate_plot or both.

It can also be used as API. There're several functions: one_hot, getlabel, flatten, convmodel3, convmodel5, test_plot and evaluate_plot, most of them will have a introduction in the beginning, you can also use main as API's example code.

Here're two line charts about the trend of loss and accuracy during training process of 3/5 lays convolutional neural network (First for 3, second for 5) . I used about 700 images of car and cat to train them. As you can see, in this amount of datasets, the 3 lays convolutional neural network perform well while overfitting occured just after several epoches in the 5 lays convolutional neural network model.

To deal with the overfitting problem, I read several papers and tutorials and find a few approaches:

• Get more training data.
• Reduce the capacity of the network.
• Add weight regularization.
• Add early stop.
• Add dropout.

I choosed to add an early stop and set the patience as 10 (since we only run 50 epoches). Also, I added a dropout in each model (the amount is 0.25 for 3 convolutional neural network and 0.5 for 5) and changed some parameters like learning rate. Here's line charts after change and it performed much better (Also first for 3, second for 5) .

After optimization, the 5 lays convolutional neural network perform better and the overfitting problem is less significant even if it still occur. Also, the accuracy arises faster and higher than 3 lays convolutional neural network. The advantages of 3 lays convolutional neural network is that it learns faster and basically no overfitting and more suitable for small dataset. So for large dataset, we should use 5 lays convolutional neural network, and for small dataset or in order to prevent overfitting, we could use 3 lays convolutional neural network.