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:

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

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