Probability is the process of finding the chance of occurrence of any event. This project focuses on finding the probability of abnormality of an Musculoskeletal parts. MURA, is the dataset that has been used for this system to detect abnormalities. There are millions of people who suffer from different musculoskeletal problems like fracture, swelling, etc. If this system is built, people don’t need to wait in a long queue for long checkup. This system helps to solve the problem of detecting abnormalities within 2 minutes. Convolutional Neural Network has been used to find the abnormal detection in a feed forward fashion. The dataset as well as the user input are preprocessed where the image is converted into standard format. This system may not provide exact result, as there might be some faults while taking an image. it helps to detect abnormalities with the use of an XRay image. Finally, accuracy test was carried out and the accuracy is 83.71%.

Extraction of an Image Dominant Color is a system for getting a major colors present in an image using a K-Means clustering algorithm. In order to generate major color of given image there involves several steps namely, conversion from BGR to RGB, resizing the image, reshaping the image. K-Means clustering algorithm is applied to the processed image. K-Means algorithm is an unsupervised learning algorithm which typically make inferences from dataset using input vectors referring to labelled. Number of cluster is provided to a K-means algorithm as a parameter. The output of this algorithm is plotted using a pie chart. The algorithm is applied to the 64*64 size image and cluster value is chosen 8 where the accuracy obtained was 95.5%. The color information retrieved from this project can be used in color based image information retrieval system, a system that extract image information from image database.

Language modeling is an essential task in natural language processing and has wider applications in downstream tasks such as speech recognition, machine translation, spelling correction, etc. Language model architectures that use word vectors to represent the vocabulary do not capture the sub-word information (i.e. morphemes) and perform poorly in case of morphologically-rich languages such as Nepali. In this project, I apply convolution to word vectors formed by concatenation of character vectors to produce feature vectors. These feature vectors capture the sub-word information of the vocabulary and are passed into an LSTM layer through a Highway network to learn a probability distribution over a set of vocabulary. The language model built in this project, achieved a perplexity score of 378.81 i.e. in each prediction the language model is equally likely to predict 379 words as the correct one.

Essays are crucial testing tools for assessing academic achievement of the students. Manual grading is still used even after the implementation of various automated grading systems. This process takes a significantly larger amount of time of the evaluator and is also a costly process. Our attempt in this project is to automate the process of essay grading and grade the essay in a similar way that a human grader would do. The project aims to build an AES using a data set of 13,000 essays from kaggle.com. We divided the essays into 6 different sets based on the context. We extracted features such as grammar errors, spelling errors, part of speech counts and length of the sentence. First, the words are tokenized in order to make it easier for the computer to understand. Random forest regression and classifier were used to learn from the features and generate the parameters for testing and validation. We used an averaged perceptron to select the best score. This project has a wide scope in different educational institutions and organizations. One of its applications can be seen in various standardized tests and large-scale examinations such as the SAT and GRE.

The emergence of audio and video data types in databases will require new information retrieval methods adapted to the specific characteristics and needs of these data types. An effective and natural way of querying a musical audio database is by humming the tune of a song. This project implements a query by humming system based on locality sensitive hashing (LSH). This method constructs an index of melodic fragments by extracting pitch vectors from a database of melodies. In retrieval, the method automatically transcribes a sung query into notes and then extracts pitch vectors similarly to the index construction. For each query pitch vector, the method searches for similar melodic fragments in the database to obtain a list of candidate melodies. The result of this project is a ranked list of songs based on the query hummed by the user. The relevancy of this project is its use in content-based audio retrieval systems.

Contrast enhancement can reveal the information of the under-exposed regions in an image. Lowlight images are not well exposed to human observation and computer vision algorithms due to their low visibility. Although, many contrast enhancement techniques have been proposed including histogram-based and power-law based, those algorithms inevitably introduce contrast under-enhancement and over-enhancement. The definition of a good enhancement result is still not well-defined. We have noticed that the image that only differs in exposures can be used as reference for enhancement algorithms. The enhancement result should keep the well-exposed regions unchanged and enhance the underexposed regions. Exposure fusion algorithm is one of the image contrast enhancement algorithms that provides an accurate contrast enhancement that can obtain the result with less contrast, less lightness distortion and more visual appealing image

Art is a skill to create unique visual experiences through composing a complex interplay between the content and style of an image. Humans have mastered the skill to create art but there lacks the algorithmic basis of this process and there exists no any artificial system with similar capabilities. Artistic Neural Style transfer is an automated system which introduces a deep-learning approach to photographic style transfer that handles a large variety of image content while faithfully transferring the reference style. This approach helps to transfer that separates style from the content of an image by considering different layers of a neural network. An artistic images of high perceptual quality is created by using neural representations to separate and recombine content and style of arbitrary images, providing a neural algorithm. The neural algorithm used in my project is an optimization technique that takes three images, a content image, a style reference image(such as artwork by a famous painter), and the input image to style and blend them together such that the input image is transformed to look like content image, but ‘painted’ in the style of the style image. The system will transform the base input image by minimizing the content and style distances (losses) with backpropagation, creating an image that matches the content of the content image and the style of the style image.

Several biometric Signature play a crucial role in human identification. Several biometric techniques are planned for private identification within the past. As signatures play a crucial role in money, business and legal transaction, secured authentication becomes a lot of and lot of crucial. A signature by a certified person is considered to be the ‘seal of approval’ and remainsthe foremost and most popular means that of authentication. Signature verification can be categorized into two types: offline and online signature verification. Offline signature verification is one amongst the foremost difficult tasks in life science and document forensics. This project tends to model associate offline author freelance signature verification task with a convolutional Siamese network. Siamese network are twin networks with shared weights, which might be trained to be told a feature area wherever similar observations are placed in proximity. This is achieved by exposing the network to combine of comparable and dissimilar observations and minimizing the Euclidian distance between similar pairs whereas at the same time increasing it between dissimilar pairs.

Road accidents are dreaded incidents, which are known to take hundreds of lives in a year in Nepal itself. People lose their lives because the help is not provided to them as soon as possible. If the reporting is automated then the lives of those that met with an accident can be saved. Thus, we propose this solution, where the helmet itself initiates a communication to report the accident using the person’s phone. The helmet has vibration sensors that are cheap yet effective and productive. As the vibration changes, the Arduino reads it and send the reading to the phone via Bluetooth module. Therefore, when the rider crashes, the spike in the sensor initiates the Bluetooth module, that triggers the application on a rider’s phone to send an alert to nearby hospitals or the selected SOS number. After testing this project, we found out that the location with the SOS message is sent automatically to the SOS number when there is a spike in the sensor value. Thus, the location is precise. The solution proposed in this project does not affect the status quo by trying to modify the motorcycle itself. In Nepal, the main means of transport is either by public bus or by motorcycle. This project can be useful to those people who tend to transport by motorcycle and want an extra safety feature that is sure to notify personnel as soon as they meet with an accident.

Online news portal and other media on the internet now produce the large amount of text, which is mostly unstructured in nature. When an individual wants to access or share particular news, it should be organized or classified in the proper class. Nepali news classification is the task of categorizing the news content into the predefined category from the training news dataset. In this project, a system has been built for categorizing the content of the news into different categories using the news article from a major Nepali language newspaper published in Nepal. This project evaluates some widely used machine learning techniques mainly Naïve Bayes for automatic Nepali News classification problem. It classifies the news by analyzing the content of the news. In the system, a self-created Nepali News Corpus with 16 different categories and total 16719 documents collected by crawling different online national news portal is used. The test showed the accuracy of 83.94% in the news classification using Naïve Bayes. In the system, the user can categorize the news based on their content by analyzing the news content from various Nepali language newspaper and clicking the categorize button.

Pneumonia being one of the most concurrent prevalent lethal diseases, is responsible for several thousand deaths every year. X-rays images are commonly used to make clinical diagnosis for pneumonia but are not as revealing as MRIs or CT scans and the diagnosis is prone to be erroneous. Trained models help to make an accurate and more reliable computer-aided diagnosis of Pneumonia based on the X-ray images of lungs. In order to evaluate the Sigmoid-function, which is a variant of Transfer Learning, a Convolutional Neural Network/ VGG19 layer is used to achieve good performance. The problems and potential enhancements with the model of Transfer Learning is also used and discussed. Pneumonia Detection application user able to input the X-ray images and able to view the result whether the patient has suffered from Pneumonia or not. It shows expected result for proper X-ray images within the scope of the system. Error messages were shown when required. Time taken was moderate for moderate sized images. Excessive delays were not encountered the final accuracy obtained was 86.388 %.

Electoral system is a kind of voting system which is also known as majority voting system where the winner of an election is the candidate that received more than half of the votes cast. Since, this project is focused on automation of the Electoral system. Here, the system is developed such that it can detect the faces and raised hand of people and generate the total counts of faces and hands detected. This project is HCI based system where the raised hand of human is detected in order to track the vote of the people. The main objective of this project is to use computer vision and an important mode of interaction i.e. hand gesture to cast the vote. In order to design the proposed system, Haar Cascades Classifier (HCC) is used to detect the faces and hands. Haar-features are good at detecting edges, also has a higher execution speed and less computations are involved so, Haar-based classifiers are applicable in the proposed system. The system was found to be 86% accurate for face detection and 71% accurate for raised hands detection. The results imply, use of normal cameras do not produce high accurate results so, higher resolution cameras or IR cameras could improve the accuracy of the system significantly

Digital images have been around us for decades now. As the quality and resolution of image have been improving, so has the increase of size of images. Several attempts have been made to improve the visual quality of lossy images without losing much of the data. Availability of large training dataset and increase in computing power has brought interest in the application of CNNs. Here, a powerful CNN is presented specific to the task of semantic image understanding to achieve higher visual quality in lossy compression. Though several JPEG encoding techniques may produce optimized image, these processes are ultimately unaware of the specific content of the image to be compressed. This method makes JPEG content-aware by designing and training a model to identify multiple semantic regions in a given image. A new CNN architecture is presented specifically to image compression, which generates a map containing semantically-salient regions so that they can be encoded at higher quality as compared to background regions

Waste management has been one of the greatest problems that we as a country have been facing. There is no proper system, which guides the waste removal trucks for effective and timely removal of wastes. This project is focused on the idea of improving and managing the proper waste removal process. The project helps in automating the waste removal process by providing the real time evaluation of waste materials present in the waste bin. The project provides a system to the user using the combination of hardware and software components. By using a combination of hardware and software, the core logic of the system is implemented to get the data from ultrasonic sensors and gas sensors from the physical dustbin to the system through internet APIs along with its timestamp and location. The project architecture can be dissected into hardware side where the technologies like ultrasonic sensors, gas sensors are connected to the main central hub called Arduino UNO, which ultimately sends the data to the system. On the software side technologies like postgresql is used to store the data in the database controlled by python-backend where all the logics are implemented. In addition, the project is focused on providing a user-friendly environment, which allows user-creation, dustbin creation and assignment along with real-time evaluation and removal notifications.

Tetraplegia, paralysis caused by illness or injury that results in the partial or total loss of use of all four limbs. People suffering from tetraplegia are not able to move and walk on their own. To perform even basic chores, they need to depend on others. The use of smart wheelchair can assist those as well as other disabled people to move around. The objective of this project is to propose and demonstrate through a prototype an eye driven wheelchair system where people are able to operate the wheelchair based on the eye movement. This proposed system can help physically handicapped people to be independent to some extent. In order to design the proposed system, the we use Haar Cascades Classifier (HCC) to detect the eye regions and the Circular Hough Transform (CHT) for Gaze Tracking. The (CHT) algorithm is used to detect circle, in which hough gradient method is used to detect the circle, after which a circle is created on each pixel of the edges of the detected circle. The region with a high concentration of pixel in the circle is determined as the center. Since the shape of an eyeball is same as a circle and the CHT algorithm has a high accuracy for detecting a circular object and their feature extraction, this algorithm is more appropriate and has been applied for gaze tracking in the proposed systems.

Many companies are ignoring the fact that they have multiple copies of same file with them, which is consuming a fair amount of space on their device. Those spaces can be reduced with the help of Data Deduplication. Data Deduplication refers to the process of eliminating the copies of repeating data. This project includes the generation of hash keys in a tree like structure to compare the hash index for each file in the given directory. It is more efficient to compare the hash key than the original file because hash keys are smaller than the original files. The binary-tree searching process keeps the keys in sorted order and traverses from root node to their leaves. The tree is recursively searched for duplicate files. If there exist no duplicate files, the tree is given a null index value. File security plays an important role in protecting the information of a user from any outsiders. The use of AES algorithm provides a secure method of file encryption. The 256-bit encryption process in implemented where a public key is used in order to encrypt the file.

In real life, there are a lot of obstacles that are present around you at all times. It is easier for someone who has the capability to see things and react accordingly but in real life scenarios, there are a large number of people that are visually impaired. This project takes the training approach to effectively detect objects that are present in the environment and warn them about the obstacles that lie there. In this project, a system has been built by effectively training hundreds of thousands of data to efficiently detect objects in real life and the intermediate result from the detection has been converted to speech form. To train the dataset, Tenserflow GPU has been used and to accurately detect the objects, You Only Look Once algorithm has been used. After that, for conversion of the result to speech, Google Text to Speech has been used. The use of machine learning has been done to effectively train models. The use of Coco Dataset model has also been done to detect 80 predefined classes. It comprises of over 1,23,000 data sets for the accurate and reliable detection of the objects. The tests on the detection show the accuracy of over 70% on the detected objects but due to the lack of computational device available, the time accuracy and frames per second have been considerably slower. Both have only an accuracy of around 20%. The application can be successfully implicated for people that are visually impaired. This algorithm has been specifically chosen as it is a state-of-the-art algorithm. Initially, ScaleInvariant Feature Transform algorithm was used which turned out to be slow and less efficient than the algorithm we currently used in the system.

Contract is a means of confirming that the agreement between two people has been established. The contract plays an important role in the trade to be done between them. Another important aspect that needs to be focused during the process of contract establishment is the privacy maintenance. Blockchain maintains the privacy to maximum extent. Thus, an individual can their contracts without having thoughts about data theft. Here, in this project entitled “Contract Management using Blockchain” I have proposed a potential new contract management system that uses the blockchain technology for secure contract establishment. The contracts are created by user and are than stored in blockchain network in the form of HEX key. It is an approach to create a digital contract system using a reliable network, security and user-friendly UI. The user can only view the system with their personal identification i.e. wallet address. This encrypted identification removes the threads of viewing the contract by unauthorized user.

The construction of first compiler for the language FORTAN (formula translator) around 1956 was whose success was not at all assured. It was figured among the largest programming projects of the time. A Compiler is a black-box program which takes source code written in some highlevel programming language and compiles it into the language that computer can understand and execute. This process is done through a number of stages. Starting with the recognition of the token through the target code generation provide a basic for communication interface between a user and a processor in a significant amount of time. It is basically a simple programming language that helps in solving different functions and modules of programming in a very simple manner. More of a programming language is difficult to understand and time consuming in terms of compilation. ProgramNepal helps in breaking the bar of conventional programming language. It is dynamically typed language because it is associated with run-time values. Due to this programmer can write a little quicker because we do not have to specified types every time.0

Diabetes is considered as one of the deadliest and chronic diseases which causes an increase in blood sugar. According to the growing morbidity in recent years, in 2040, the world’s diabetic patients will reach 642 million, which means that one in ten adults in the future has diabetes. There is no doubt that this alarming figure needs great attention. The remarkable advances in biotechnology and health sciences have led to a significant production of data, such as high throughput genetic data and clinical information, generated from large Electronic Health Records (EHRs). To this end, application of machine learning and data mining methods is vital and indispensable in efforts to transform intelligently all available information into valuable knowledge. The motive of this study is to design a model which can predict the likelihood of diabetes in patients with maximum accuracy. This project implements four linear model and one deep learning model: Logistic regression Naïve Bayes, Support Vector Machine, K-Nearest Neighbors and Multilayer Perceptron Neural network to investigate performance on diabetes disease dataset obtained from UCI data repository. For the training purpose, 768 datasets have been used, among which 500 are non-diabetic and 268 are diabetic. The performances of all the algorithms are evaluated on the basis of accuracy. In addition to the comparison of the algorithms, each algorithm has been integrated into a prediction engine and exposed over an API. As the results show, Multi-Layer Perceptron Neural Network outperforms other linear models however K-Nearest Neighbors gives identical results with less computing overhead. Performance improvements could also be achieved by using complex deep learning methods

There are different learning algorithms to train a bot while playing a game. The main objective of this project is to implement reinforcement learning and genetic evolution algorithm to train a bot. The genetic evolution falls under hybrid method according to their goal, focus and component methodologies. In this project we are trying implement both algorithm and implement it in a custom endless runner game “Flying Deer”. In the case of genetic evolution, a generation of bots will be created and the bot plays the game. Certain fitness function will be given according to which the bot plays the game and the fittest among the generation get mutated and replaced in such a way that the bot will play the game endlessly. Similarly, in the case of reinforcement learning a single bot plays the game. The deer will be rewarded and punished with every iteration of the game. On the basis of the reward, the brain of the bot is trained to avoid the obstacles and play the game endlessly. Both the learning algorithm will be implemented which will successfully be able to learn about the game environment and play the game continuously

Parts of Speech Tagger for Nepali Text Using SVM is an application that assigns the appropriate parts of speech like noun, pronoun, verb, adverb, adjective etc. and other lexical tags to each words written in Nepali language based on its definition as well as context. The parts of speech tagger is build using the supervise machine leaning algorithm namely Support Vector Machine. The model uses 14 million Nepali words and corpus consists of written text from 15 different genre with 2000 words each published between 1990 and 1992 and the texts from a wide range of sources such as internet webs, newspapers or books. And, the model is trained with 80,000 lemmatized words collected from the Nepali National Monolingual Written Corpus. The Parts of Speech tagger for Nepali text has wide range of scope in research and NLP applications such as machine translation, speech recognition, speech synthesis, grammar checker, information retrieval and extraction. Nepali is morphologically rich language and one has to consider many features to build the language model. The SVM based POS tagger construct the feature vectors for each word in input and classify the word into one of the two classes (One Vs Rest). The performance analysis includes different components such as known words, unknown words and size of the training data. The average accuracy obtained for lemmatized text and unprocessed raw text is 88% and 72% respectively

With the growth of internet, the online news publishing site have increased tremendously. The major news houses have shifted their focus towards online publication rather than the prints. With the evolution of online news publication, anyone can publish anything without caring about the impact that it could make it to the society. The maximum number of news published digitally are negative and have been feeding negativity among the general public. This project “Sentiment Analysis of Nepali news headlines” is carried out so that the reader can read the news of their preference i.e. positive, negative, or neutral as per their choice. The system automatically analyses the sentiment of the news using machine learning and places them in one of the three categories. The news from major news publishing sites can be grouped using this project.

Music genres are categories that can be used to identify and arrange growing amount of music emerging at present. Conventionally, music genre labeling was done manually. However, due to the advent of new techniques and growing number of researches in Music Information Retrieval, some form of automation has been seen in the field of music genre categorization. In this project, a stepping stone for automatic music genre categorization of vast number of music files available in digital form online or offline has been developed. Out of the various techniques of music genre recognition, content-based technique is used to automatically label the particular user uploaded song into one of five distinct genres: Classical, Pop, Metal, Jazz, and Blues. Digital signal processing techniques of Fast Fourier Transform and Mel-Frequency Cepstral Coefficients have been used to generate feature values of feature vector, which is then fed into the classifier developed using Support Vector Machine, in order to classify user input song. The training and testing of the system has been performed successfully obtaining an accuracy of 74.0%, which is significant in the field of music analysis. Training has been carried out using 80 music clips of each of the five genres, and testing, using 20 music clips of each of the five genres. GTZAN Music Dataset, a popular Western music dataset prepared for music analysis, has been used during training and testing; hence, this system works well only with Western music files. The system has been implemented with two graphical user interfaces: one for admin (training and testing) and another for user (uploading music file and finding genre) using Python programming language and Flask framework

Image captioning is a standout amongst the latest issue that got enthusiasm of PC vision network and Natural Language Processing people group. Image Captioning is the method where programmed portrayals are created for a picture. Consequently, portraying the substance of a picture is a basic issue in man-made consciousness that associates PC vision and normal language handling. Dataset is prepared to amplify the probability of the objective portrayal sentence given the preparation image. This task tends to the issue by utilizing a profound neural system model. The model would utilize Convolution neural systems to examine the image information and Recurrent neural systems for learning sentences/inscriptions for image. Flickr8K datasets is utilized to prepare the model. TensorFlow system is utilized to exhibit continuous relevance and its improvement. Lastly, the trained neural network helps generate the caption according to the image uploaded by the user.