This project focuses on semantic image segmentation, specifically targeting images of cats and dogs. Semantic image segmentation involves classifying each pixel in an image into a specific class, allowing for detailed understanding and analysis of image contents. In this project, I implement two popular architectures, U-Net and SegNet, for semantic segmentation and compared their performance.
The inspiration for this project originated from François Chollet's image segmentation classifier as presented in the book "Deep Learning with Python". Expanding upon this concept, I aimed to employ architectures like to U-Net and SegNet for image segmentation, intending to compare their performance.
- Symmetrical architecture with equal downsampling and upsampling layers.
- Skip connections preserve spatial information.
- Encoder-decoder architecture.
- Uses max-pooling indices for upsampling, reducing complexity.
- Designed for efficiency, suitable for real-time and resource-constrained scenarios.
After 30 epochs, the U-Net-like architecture achieved a training loss of 0.2429 and a validation loss of 0.3614. On the other hand, the SegNet-like architecture had a training loss of 0.4011 and a validation loss of 0.4176.
|
U-Net
|
SegNet
|
- Increase the number of images and their corresponding annotations.
- Deepen the network by adding more layers and narrow it by reducing the number of neurons.
- Introduce batch normalization layers after convolutional layers to enhance performance.
- http://www.robots.ox.ac.uk/~vgg/data/pets/data/images.tar.gz
- http://www.robots.ox.ac.uk/~vgg/data/pets/data/annotations.tar.gz
This project is licensed under the Raza Mehar License. See the LICENSE.md file for details.
For any questions or clarifications, please contact Raza Mehar at [[email protected]].