Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/amansharma2910/deeplearning-sceneclassificationcnn

In this project, originally hosted on Kaggle, I have worked on a Deep Learning model based on the powerful CNN algorithm, capable of detecting the scene in an input image.
https://github.com/amansharma2910/deeplearning-sceneclassificationcnn

Last synced: 30 days ago
JSON representation

In this project, originally hosted on Kaggle, I have worked on a Deep Learning model based on the powerful CNN algorithm, capable of detecting the scene in an input image.

Awesome Lists containing this project

README 

        
Scene Classification Using CNN


 

In this project, originally hosted on Kaggle, I have worked on a Deep Learning model based on the powerful CNN algorithm, capable of detecting the scene in an input image. 



\***[Link to Kaggle Notebook](https://www.kaggle.com/bravehart101/sceneclassification-cnn-pytorch)**



Project Objectives



The following are the project objectives-

* Creating a Google Images-like multi-class scene classifier using the CNN algorithm.

* Working on different CNN architectures and comparing their performance.

* Comparing performance gain (or loss) while working with different deep learning techniques.  



Understanding the Problem Statement



The dataset that I have used for this project is the Intel Image Classification dataset. 



This is a multi-class classification problem where, given a picture, our model processes the image then classifies it as one of the six possible classes- 

 * buildings: 0

 * mountain: 1 

 * street: 2

 * forest: 3 

 * sea: 4 

 * glacier: 5



As per the data source, each image is a 150 × 150 pixel, 3-channeled (RGB) image, with around 14k images in the training set and 3k images in the testing set.



Observations


For the scene classifier, I worked with two CNN architectures— A custom ResNet9 model, and a ResNeXt9 model based on the ResNet9 model where I replaced the residual blocks with ResNeXt blocks of cardinality 4.



The ResNet9 model had around 7 million learnable parameters, while on the other hand, the ResNeXt9 model had around 22 million learnable parameters (around 3 times more). Here are the observations after training and testing the model-



* After training with 10 epochs, the ResNet9 model an accuracy of somewhere about 83% on the test dataset. At the same time, it took around 15 epochs for the ResNeXt9 model to achieve 81% accuracy on the test set.

* Increasing the architectural complexity reduces the speed of training. Also, more learnable parameters doesn't guarantee a performance boost. Rather, making the architecture more complex can result in overfitting, thus causing the classifier system to lose generalization capabilities which might cause it to perform poorly on test/real-world data.

* Normalizing the data can significantly reduce learning time while also showing a boost in accuracy. This observation was derived with the ResNet9 model which showed around 5% accuracy after normalizing both the training and testing dataset.



Key Highlights



 This is a complete, end-to-end project, where, in order to achieve the optimum performance I have implemented different Deep Learning techniques, namely-

 

 * **data normalization** 

 * **data augmentation**

 * **batch normalization**

 * **weight-decay-prevention** 

 * **learning rate scheduling**  

 

 Along with above mentioned techniques, I have also tried using different CNN architectures, comparing the model performance for each architecture.

 

 \*Note- *Each separate jupyter notebooks contains a different model architecture along with different transformation and optimization techniques. This is an ongoing project so I'll keep adding more notebooks with different custom CNN architectures in the future.*