Browsing by Subject "Deep Learning"
Now showing 1 - 5 of 5
Results Per Page
Sort Options
Item 3D RECONSTRUCTION BASED VIRTUAL TOUR(I.O.E. Pulchowk Campus, 2023-04) KOJU, BISHAD; JYAKHWA, GAURAV; NYOUPANE, KRITI; MANANDHAR, LUNANumber of research and several methods have been proposed for 3D reconstruction from 2D images. The first is a triangulation approach based on determining the same points in images taken from different angles to approximate a point cloud in 3D space and then reconstructing the mesh. This is purely a computation-based approach. Another approach is to redefine 3D reconstruction problems as recognition problems and use the existing knowledge about 3D space and projection to reconstruct, much like how humans do. This knowledge is approximated using deep learning models. However, in these approaches, the mesh reconstruction part is extremely expensive. This cost can be reduced by trying to reconstruct the view rather than trying to reconstruct the mesh. Neural Radiance Field (NeRF) has been used to generate novel views. NeRF represents a scene using a fullyconnected deep network, whose input is a spatial location and viewing direction and output is the volume density and view-dependent emitted radiance at that spatial location. We synthesize views by querying 5D coordinates along camera rays and use classic volume rendering techniques to project the output colors and densities into an image. In this project, we have used the latter approach.Item An Assessment for Predicting Stroke Patients using Bidirectional LSTM(Pulchowk Campus, 2019-11) Puri, AjayaStroke is the medical condition when the supply of blood to the brain is either interrupted or reduced for very certain duration of time. When this happens, the brain does not get enough oxygen or nutrients, and brain cells start to die. This thesis presents the development and evaluation of a machine learning model using deep learning techniques. The improved, memory based Bidirectional recurrent neural called Bidirectional Long short-term memory (BLSTM RNN) is used for the research work. The model thus developed predict whether a patient will experience stroke or not based on a time series input data computation. A 3-layer architecture having single BLSTM unit, Adam as model optimizer and dropout regularization of 0.42 achieves accuracy of 91%. The model is developed by processing patient time series information which includes demographic and medical historical data. It includes age, gender, hypertension, heart diseases, and altogether ten biometric information. This work contributes for decision support for individuals and medical persons on their future stroke possibility.Item “INFORMATION EXTRACTION FROM UNSTRUCTURED DATA”(I.O.E. Pulchowk Campus, 2023-04-30) LAMMICHHANE, AAYUSH; NEUPANE, AAYUSH; PAUDEL, ANKIT; LAMSAL, ASHISHIn today’s digital age, the digitization of paper documents like invoices and receipts has taken on more significance. Nevertheless, manually entering data from these papers can take a lot of time and be prone to mistakes, which causes inefficiencies and drives up expenses for enterprises. To solve this issue, we created a software platform that automates the process of collecting important data from scanned documents using deep learning technology, more specifically the LayoutLM architecture. Users can upload their scanned papers in bulk to our platform and choose which fields, including date, merchant name, and total amount, they want to extract. The technology is scalable and can manage high document volumes while preserving precision and effectiveness.The user-friendly interface makes it easy for users to upload and extract information from their scanned documents. Our platform offers significant benefits, including increased efficiency, accuracy, and cost savings, and has the potential to transform the way businesses handle physical documents. In this project, we will provide an overview of our software platform, including the technology behind it, its key features, and its potential applications.Item STOCK PRICE PREDICTION USING DEEP NEURAL NETWORK(Pulchowk Campus, 2021-08) PHULARA, BASANT RAJForecasting the financial market is one of the practical problems in the economic field. The noisy and the volatility are the two characteristic that hinders the timely prediction of the stock future price. In order to further resolve the drawbacks of the existing models in dealing with non-stationary and non-linear characteristics of high frequency financial time series data, this research work proposes the Wavelet transform based data preprocessing and developing the LSTM-attention model including the human sentiment for stock price prediction. The financial time series is smoothened by Wavelet transform, LSTM and attention mechanism is used to extract and train its features. Also the impact of human sentiment is investigated by adding the sentiment polarity score to historical dataset. The results of the proposed model are compared with the other two models, including LSTM and GRU on four different stocks ADBL, NIB, NABIL and SCB datasets. The Performance of the different models is evaluated based on RMSE, MAE, coefficient of determination R2, and MDA. The results from the experiments on all the stock datasets shows that RMSE, and MAE is less than 2.5 and 2.2 respectively and R2 is greater than 0.94 and MDA greater than 0.79. The results show that the proposed model along with the addition of human sentiment outperforms other similar models.Item Video Summarization using Spatio-Temporal Features by Detecting Representative Content based on Supervised Deep Learning(Pulchowk Campus, 2021-08) Sah, Ramesh KumarVideo Summarization is the approach to generate the compact version of video keeping relevant content intact and eliminating redundancy. In this work, a frame- work has been proposed which makes use of the spatial and temporal features with self attention from the video sequences to identify the representative con- tent by generating temporal proposals and supervised learning from the data manually created by humans or users. Existing Supervised methods don't deal with the temporal interest and its consistency. For that temporal uniformity is also necessary which can be addressed by predicting the temporal proposals of the video segment. The proposed work treats it as temporal action detection which predicts importance score and location of the segments simultaneously by developing the anchor based method which generates anchors of varying lengths to identify interesting proposals. Moreover the extensive quantitative and qualitative analysis on TVSumm and SumMe datasets augmented with OVP and YouTube datasets justify the e ectiveness of the method.