Skip to content

Deep Convolutional Encoder-Decoder Architecture implemented along with max-pooling indices for pixel-wise semantic segmentation using CamVid dataset.

License

Notifications You must be signed in to change notification settings

chandnii7/Image-Segmentation

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

18 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

SegNet : Deep Convolutional Encoder-Decoder Architecture for Pixel-wise Image Segmentation

This Project is implemented based on the following research paper: View Research Paper
Program was implemented using Python, TenserFlow, Keras and OpenCV. Refer the report for further implementation details: View Report
CamVid dataset is used for training. Download CamVid dataset from provided drive link into CamVid folder: Download CamVid
Trained Model weights can be downloaded from provided drive link into Model folder: Download Model



Network Architecture

Overview:

  • Contains encoder network and corresponding decoder network which will consist of a hierarchy of decoders one corresponding to each encoder
  • Encoder network has 13 convolutional layers
  • Decoder network has 13 layers corresponding to each encoder
  • Non-linear upsampling in decoder using pooling indices from max-pooling step of the corresponding encoder for accurate boundary localization
  • Upsampling maps are convolved with trainable filters to produce dense feature maps in decoder because upsampling maps are spared
  • Followed by final pixel-wise classification layer using Softmax
  • Dataset augmentation with albumentations


Results:

  1. Comparing Ground Truth and Predictions of Trained Model:


  1. Evaluation of SegNet on Train, Validation, and Test Data:


  1. Precision, Recall, F1-score and Support:


About

Deep Convolutional Encoder-Decoder Architecture implemented along with max-pooling indices for pixel-wise semantic segmentation using CamVid dataset.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published