[Model and README Enhancement] Glass Bangle Defects Detection

Glass-Bangle-Defect-Detection/README.md → Glass Bangle Defects Detection using DL/README.md

@@ -15,8 +15,11 @@ The data images(about 1080 in number) were especially large (3000X3000 pixels).
 - Normalized the data images using empirical values of mean and standard deviation.
 - Finally all images were converted into batches and loaded into the pytorch model
 
+- For CNN with Attention, we took images as 400 X 400, normalized them, and then flipped the colour intensity to highlight contours better.
+
 ## The Model
 So the model I have used here is here is ResNet and GoogleNet, commom industry standard models developed by researchers at Microsoft and Google respectively.
+
 #### ResNet
 <img src="./Images/Residual-Block.PNG" width="800" title="CNN Model">
 
@@ -28,26 +31,36 @@ So the model I have used here is here is ResNet and GoogleNet, commom industry s
 #### Custom-Model
 Also I have made a custom model using one convolutional layer and 2 fully connected layers in order to display the internal working of the CNN-architecture based Neural Networks.
 
+<img src="./Images/Residual-Block.PNG" width="800" title="CNN Model">
+
+#### CNN
+
+A Deep neural network model which has feature selection and feature training components.
+
+<img src="https://github.com/Arihant-Bhandari/DL-Simplified/blob/enhance_bangle_defect/Glass%20Bangle%20Defects%20Detection%20using%20DL/Images/CNN-Keras.png">
+
+#### CNN with Attention
+
+A Deep neural network model which has feature selection and feature training components, with an overhead attention mechanism to introduce probability distributions into training.
+
+<img src="https://github.com/Arihant-Bhandari/DL-Simplified/blob/enhance_bangle_defect/Glass%20Bangle%20Defects%20Detection%20using%20DL/Images/CNN-Keras-Attention.png">
+
 ## Training and Testing of the Model
-ResNet and GoogleNet took approximately 19 minutes each to be fine-tuned and generate accurate results.And the custom model took approximately 30 minutes of computational time. All models were trained on GoogleColab using 12 GB free GPU provided by Colab.
+ResNet and GoogleNet took approximately 19 minutes each to be fine-tuned and generate accurate results.And the custom model took approximately 30 minutes of computational time. All models were trained on GoogleColab using 12 GB free GPU provided by Colab. CNN-Keras and Attention were trained on Kaggle with GPU-T4 boosting.
 
 | Models | Best Train-Accuracy | Best Validation-Accuracy | Computation-Time| 
 |--------|---------------------|--------------------------|-----------------|
 | ResNet-18 | 86.21% | 89.8618% | 19m 31s |
 | GoogleNet | 74.62%  | 77.8802% | 19m 42s | 
-| Custom Model |53.548% | 53.548 % | 30 minutes | 
-
+| Custom Model | 53.548% | 53.548 % | 30 minutes | 
+| CNN (Keras) | 89.9% | 81.4% | 2.5 minutes | 
+| CNN (Keras-Attention) | 99.32% | 88.27% | 1.67 minutes | 
 
 ## Conclusion
 
-So the best model selected is going to be ResNet which is a very good model for Image Recognition. For more information on ResNet, read https://arxiv.org/abs/1512.03385
+So the best model selected is going to be ResNet which is a very good model for Image Recognition, and another model to be considered is CNN with Attention mechanism with comes just as close.
 
-## Courses I followed to Build the Model 
-- Michigan's Deep Learning for Computer Vision
-(https://www.youtube.com/playlist?list=PL5-TkQAfAZFbzxjBHtzdVCWE0Zbhomg7r)
-- Pytorch Tutorials
-(https://www.youtube.com/playlist?list=PLqnslRFeH2UrcDBWF5mfPGpqQDSta6VK4)
+## Your signature
+Original Contribution: Aryan Gupta 
 
-## Developer Details
-Developed By Aryan Gupta 
-(Electrical Engineering Student, IIT Roorkee)
+README and Model enhanced by: Arihant Bhandari [https://github.com/Arihant-Bhandari]