In a student's life notes are very useful during exam time. But after exams are over, the gallery section is filled with random scattered images which take's extra space. Note separator is programmed to detect notes images automatically, to ease the process of deletion. The Project was merged under DSC IIT Patna's NoteSeparator
Dataset that is used to train the model can be found at link. The above dataset contains total of 1500 images (aprox.) collected with the help of IIT Patna Student Fraternity.
- 850 notes images (aprox.).
- 650 non-notes images (aprox). Size of each image later reduced to 150x150.
- TensorFlow (Keras)
- Numpy
- Pandas
- Matplotlib
- OS
- PIL (Python Image Library)
The model showed maximum accuracy when the learning rate was reduced to 0.001. It consist of 5 convolution layers. Number of nuerons in each layer from top to bottom were 16,32,64,64,64. With RELU used as activation function in each layer followed by a maxpooling layer and a dropout of 0.2(initial 2 layers).
Results were then flattened to be fed into DNN.
There was 1 Dense fully connected layer consisting of 512 nuerons .
Lastly 1 more dense layer consisting of single nueron for the output with sigmoid as activation function. Since dataset was small techniques like Data Agumentation had to be used.
- Test acc. 75% (Lower owing to generic - less processed - dataset)
- Training acc. 88.68%
The model could be used with a mobile application, though with current accuracy levels, it would be hard not get sweaty fingers while tapping the 'delete selected' button.
SInce, this was my first fling with ML/Dl, this project helped me to learn the following
- CNN and DNN
- Agumentation technique
- Optimizers
- Using callbacks
- Confusion matrix
- Use of PIL (python imaging library)