Segnet-Transfer-Learning

This project is based on the Python language, the Keras deep learning framework, migration learning techniques and the tensorboard module to build a SegNet neural network, documenting the training process.

Installation

tensorflow-gpu==1.13.1

keras==2.1.5

Run project

You can use our trained deep learning models for semantic segmentation prediction of city street images.

1.Place your prepared street view images in the img folder.

2.Change the image read path in <predict.py>.

3.Run <predict.py> for semantic segmentation of street view image.

4.The results are saved in the imgout folder.

Download the cityscapes dataset

The Cityscapes is an open data set that can be downloaded from the official website. (https://www.cityscapes-dataset.com/)

Train your own neural network

1.Prepare the cityscapes training dataset and modify the path to the dataset in the train file.

2.Open the <train.py> and change the training parameters, which I have as follows:

batch_size = 2

Transfer training:

epoch = 20

optimizer = Adam(lr=1e-3)

Global training:

epoch = 40

optimizer = Adam(lr=1e-4)

3.Run the <train.py>.

4.The output neural network weights will be saved in the logs folder.

Neural Network Accuracy

Use tensorboard record the loss function and accuracy of the training process.

Transfer training stage stored in Tensorboard1 folder.

Global training stage stored in Tensorboard2 folder.

Use command line cmd and tensorboard --logidr=Tensorboard1 tensorboard --logidr=Tensorboard2 get url to open chrome visualisation of the training process.

Following is an example of a visual file of our team training process:

Transfer training stage Training set accuracy

Transfer training stage Validation set accuracy

Global training stage Training set accuracy