Internship Experience

Internship Experience:

Software Developer:

Working with C#, SQL programming languages at Visual studio, server management studio using Dot Net & ASP.Net and Gained knowledge on .Net, ASP.Net and SQL client-based projects

Time Period: Jul '21 - Sep '21

Company: INFOLOG SOLUTIONS Pvt Ltd, Bangalore
Programming Language: ASP.Net, C# and SQL
Technology: Software Development

Deep Learning Engineer:

My work includes an Internship related to Deep Learning and Computer vision at Codegnan IT solutions located in Vijayawada.

Time Period: January 06, 2020, to March 20, 2020

Company: Codegnan IT Solutions, Vijayawada

Project 1:

Face Recognition for Attendance based system: Working on face Recognition for Attendance System with deep learning and convolutional neural networks using Tensorflow, Keras and OpenCV

1. Create Database: The first step in the Attendance System is to create a database of faces that we will be using. Different subjects are taken and a camera is used to detect faces and record their frontal face. The number of frames to be taken for consideration can be modified to our accuracy. These images are then stored in the database with the Registration ID. Using the OpenCV technique we will try to take the Images from the webcam by applying a Face detection cascade file.

2. Face Detection: Face detection is the process of automatically locating faces in a photograph and localizing them by drawing a bounding box around their extent. For face detection, we are using a cascade .xml file. It is the pre-trained file that we used for face detection. And as mentioned above in creating a database the face detection process applied. So that the faces will detect the only person's faces in terms of either Color or Gray. Extract face from images, crop face, and store as an image in a separate folder with the image name as person name

3. Training Faces: After taking the dataset try to train them and create a model for recognition. There are many types of models used in the real world. Those are VGG, OpenFace, Facenet, Deep Face, etc. Here we are using the VGG16 model for training. VGG model is one of the Neural Network architectures containing 16 trained layers. Once the model is built we can set the layers weights to values trained on a larger dataset. For the VGG16 model, we are using pre-trained weights and by adding weights to the model we can interact with the dataset to train. Once the model has been fully trained (architecture and weight value) we can use it to identify faces using the predict method. As the VGG weights have been trained on square/centred face images we test it on an image that has been manually cropped.

4. Face Recognition: For face Recognition, Earlier we stored each cropped face image in the corresponding folder, walk through each image, load the image from Keras in-built functions. Now we can recognize any face in the image if we get embeddings for the face with the help of the VGG model and feed it into to classifier then get a person's name. OpenCV draws a rectangle box around the face and writes a person's name for each face in the image.

5. Post Preprocessing: After the xlsx sheet is applied to the system those data will be uploaded to corresponded authorities. Using matplotib, the email package we did the process of sending the file and text.

Project 2:

To control the media player using Face Detection with Javascript:

1. We worked on face detection with a haar cascade file named “haar_cascade_frontface.xml”. This project tells that when the face is present on the server then only the video will play on the server. If the face is not on the screen the video automatically pauses. Implemented this by using Javascript with the help of an OpenCV file named opencv.js. The video will pause using the video tags.

Project 3:

To control the media player with Hand Detection using OpenCV:

1. For the Hand detection, we created a haar cascade with the help of positive and negative images and train those images to recognize the Hand. But it is not accurately showing the best result when we created it. So, we try to fist the haar cascade files for controlling the media player. Using that file when the first is detected it needs to give the binary value as “1” and no hand detected it needs to give the value as “0”. Those backend data will turn t the front end named Javascript. Again using the Binary value the video tages describe the video to pause or play.

Project 4:

Multi-class classification using TensorFlow and Bert Model:

1. The bert model is built on the top of the techniques named “seq2seq” models and transformers.

2. Seq2seq: The seq2seq model is a network that converts a given sequence of words into a different sequence and is capable of relating the words that seem more important. Ex: LSTM network.

3. Transformers: The transformer architecture is also responsible for transforming a sequence into another, but without depending on any Recurrent Networks such as LSTMs or GRUs.

4. To implement the BERT model we need certain requirements to proceed:

Data preprocessing:

Normalizing the text by converting all whitespace characters to spaces and casing the alphabets based on the type of model being used(Cased or Uncased).
Tokenizing the text or splitting the sentence into words and splitting all punctuation characters from the text.

Reference Links: a. https://www.youtube.com/watch?v=p-Jf_FRqjXQ

b. https://www.youtube.com/watch?v=76zoEO29ZdQ

Research Internship:

I had completed an Internship under the professor's supervision named Prof.Ramamurthy Garimella located Mahindra Ecole Centrale, Hyderabad and worked in Hopfield Associative memories and Deep Learning, fields. We submitted two Research papers on IJCNN and IntellySis 2020.

Time Period: December 10, 2019, to February 10, 2020

The paper named "Deep Neural Networks: Incremental Learning" has been reviewed and accepted for presentation at the Intelligent Systems Conference (IntelliSys) 2020

Project 1:

1. Hopfield Associative memories: I worked in the areas of Hopfield Associative Memories and Deep Learning architectures such as Convolutional Neural Networks, and Autoencoders along with a team. Did Neural Networks and Deep Learning Applications in both Serial and Parallel mode project named "Parallel Stacked Hopfield Associative Memories: Deep Learning". The Hopfield neural network belongs to the field of Artificial Neural Networks and it is a form of Recurrent neural network, that provides a model to understand human memory. In this project, the Hopfield neural network has Symmetric weights with no self- connections. During the training of the Hopfield neural network, the weights will be uploaded having the state matrix (input matrix) in the range of {+1, -1}, multiplied with weights, and subtracted by the Threshold value. Check Hopfield neural network both in Parallel and serial mode with the initial values given above it comes down as Stable stage and Converged. Next, apply these parallel mode applications will be applied for Image Reproduction, from the architecture of the Hopfield neural network the Image will be reproduced. At last, using the PYTHON the architects of Hopfield neural network and Image Reproduction one of the Deep Learning application will be proved mathematically and practically.

Research Paper: https://www.researchgate.net/publication/348838879_1-D2-D3-D_Hopfield_Associative_Memories

Project 2:

2. Deep Neural Networks: Incremental Learning: The project of Incremental Learning is addressed. Based on the idea of extracting features incrementally using AutoEncoders, CNNs, Deep Learning architectures are proposed. We worked on both MNIST and CIFAR 10 Datasets. Created 2 models with 6 Convolution, 4 Max pooling layers without having flattering, Put the trained autoencoder in parallel, and feed them to fully connected layers. The first fully connected layer comprises 256 neurons with Relu as the activation function the output of which is connected to the second fully connected layer with 128 neurons by using the activation as Relu. Then a dropout of 0.2 is used. The final fully connected layer is included with 2 neurons as the output with Softmax as the activation function. In the final merged model, we used loss function as the binary cross-entropy with optimizer Adam. Finally, we worked on 7 architectures and train them using the datasets given as input. Check the accuracy and compare it with the results.

Research Paper: https://www.academia.edu/77136583/Incremental_Learning_Deep_Neural_Networks

Certificate: Click here

Letter Of Recommendation: Click here

Reference: Click here and Click here