Smoke detection via semantic segmentation using Baseline U-Net model and image augmentation in Keras
This repo is a partial implementation from Kaggle
The main purpose of this use-case is to detect smoke in any background. The smoke can also have variations regarding its source, color, environment etc. We should be able to semantically segment smoke to analyze it's various features like color, intensity, duration of ejection of smoke (from video feed), etc.
The master branch has implementation of U-Net, however another implemetation using LinkNet is provided in different branch.
The U-Net is a convolutional neural network that was developed for biomedical image segmentation at the Computer Science Department of the University of Freiburg, Germany. The network is based on the fully convolutional network and its architecture was modified and extended to work with fewer training images and to yield more precise segmentations.
Image source: Computer Science Department of the University of Freiburg, Germany
In Kaggle Airbus ship detection challenge, Kevin Mader has used this model starting with filter size 8, for detection of ships from 768x768 image. However I have used it on 'smoke images' obtained from Google search and resized them to 256x256.
- The dataset has around 400 images, adding more images to dataset can improve the accuracy
- Proper annotation of smoke also affects the prediction of the model, maybe the annotations done in the dataset can be improved and it will surely improve accuracy
- Impelementation of the original U-Net model can also improve accuracy
