http://210.212.227.212:8080/xmlui/handle/123456789/161 Sun, 17 May 2026 00:07:20 GMT 2026-05-17T00:07:20Z http://localhost:8080/xmlui/bitstream/id/3ad6754b-3899-444f-b659-78e7a669ed03/close-up-handwritten-whiteboard-year-168763957.jpg http://210.212.227.212:8080/xmlui/handle/123456789/161 http://210.212.227.212:8080/xmlui/handle/123456789/273 TEXT EXTRACTION FROM LOW QUALITY IMAGES USING DEEP LEARNING TECHNIQUES Asha, Simon; Aneesh, G Nath Today whatever look can see presence of digital image.Every day may groups like doctors, engineers, and students etc. release many images for their dif ferent needs.These images may contain both textual and non textual data. Text present in such images can provide meaningful information for content based repossession and many applications of computer vision like image un der standing, reading number plates of moving vehicles and context retrieval for the investigation purposes.Text information found in these images may very different in fond size, style, alignment and orientation. In this way, it is very difficult to identify items with different characteristics and even more difficult to retrieve text if image is blurred or low resolution.There are many existing techniques for recover textual information from clear image. So in this paper, introduces a method to recover text from a blurred or low reso lution image. The proposed technology is comprised of three main steps.(a) The deblurring process is applied to recover the clear image. (b) Extract text from images using text localization, segmentation and binarization tech niques (c) Evaluation of proposed model. Input image debluring achieved by using super resolution convolution network. Text extraction can be achieved by using text extraction network. Here, use text detection to identify the text region on input image after that can find the exact position of text by using text localization and text segmentation separates the text from its background. Extensive experiments have been conducted on a synthetically generated dataset. Experimental results and analysis show that this system has better performance in terms of quantitative evaluation. Thu, 01 Sep 2022 00:00:00 GMT http://210.212.227.212:8080/xmlui/handle/123456789/273 2022-09-01T00:00:00Z http://210.212.227.212:8080/xmlui/handle/123456789/272 A Performance-Sensitive Malware Detection System Using Deep Learning on Mobile Devices Athira, Joshi; Ansamma, John Concerns about the threat of piracy and the spread of viruses have been raised as a result of the rise in popularity of Android applications. Through applications, malware is frequently propagated in the mobile environment. People frequently grant applications rights without fully understanding their purpose. Android devices are used by almost 2.8 billion people worldwide. Therefore, the requirement for a reliable Android malware detection system is unavoidable. In order to provide a real-time and responsive detection environment for Android mobile devices, this paper proposes MobiTive, an efficient Android malware detection system. A more practical and secure application solution is MobiTive, which checks and monitors APK files before installation. MobiTive is concentrating on mobile and server-side platforms. APK files are mined for API calls and Manifest files of both malware and benign apps, and this raw data is then fed into convolutional neural networks. After being trained independently, networks including GRU, BI-GRU, STACKED GRU, and LSTM were shown to be of good performance. Its shown that BI-GRU have the highest performance. As quantization is done for mobile device adaptation, network model size varies. Deep Learning model must be converted to Tensorflow-lite model due to limited computing power and resource availability in order to function well on the mobile platform. MobiTive has been evaluated for effectiveness on a variety of mobile devices, and it achieves detection accuracy of 90.26% percent. Mon, 01 Aug 2022 00:00:00 GMT http://210.212.227.212:8080/xmlui/handle/123456789/272 2022-08-01T00:00:00Z http://210.212.227.212:8080/xmlui/handle/123456789/271 A COURSE RECOMMENDATION SYSTEM BASED ON LEARNER CAPACITY USING RESTRICTED BOLTZMANN MACHINE Dhanya, M; Thushara, A The education field has undergone a rapid transformation with the advent of online courses, that aid in skill acquisition and life long learning. However, in the present scenario, the major limitation of these courses is the high dropout rate. This is due to the poor course recommendation and the low learner in teraction associated with online courses. In order to reduce the dropout rate, course recommendation must be given based on the learning capacity of stu dents. In this weak learner interaction scenario, cognitive diagnosis, which is a psychometric technique used to discover the proficiency level of students in specific knowledge components could not be used. Instead, a variant of this technique called Multi-dimensional Item Response Theory (MIRT) can be in corporated in course recommendation systems to suitably represent learner’s learning state by obtaining implicit response on the followed course. In this work, course recommendation based on learner capacity is implemented by integrating MIRT into collaborative filtering, which is implemented using Restricted Boltzmann Machine. The Open University Analytics Dataset is used as the base experimental data. Experimental results and analysis show that this course recommendation system has better performance in terms of different quantitative metrics like precision and mean absolute error. Fri, 01 Jul 2022 00:00:00 GMT http://210.212.227.212:8080/xmlui/handle/123456789/271 2022-07-01T00:00:00Z http://210.212.227.212:8080/xmlui/handle/123456789/270 INTRUSION DETECTION SYSTEM USING OPENPOSE Divya, C P; Dimple A, Shahjahan Intrusion Detection System(IDS) is a monitoring system that identify un usual activity and send out alerts when it does. It is a core part of site’s safety and security strategy. A computer vision task that requires recogniz ing, associating, and tracking semantic key points is human posture estima tion and tracking. Human Action Recognition (HAR) methods are gaining importance, with the emerging advancements in computer vision and pat tern recognition. There are various methods present for action recognition and each technique has its advantages and disadvantages. Despite being a lot of research work, action recognition is still a challenging and complex task. The precision of posture estimate has been limited because of the significant processing resources required for semantic keypoint tracking in live video data. Human Pose Estimation is a core problem for understanding people in videos and images. This project presents an effective method for improving existing security systems and to further automate the threat prevention and protection procedure. This method identifies human wall climbing activity, that occur within the video range and detects unlawful or suspicious activi ties, alert if detected. This work is divided into two parts; multi person pose estimation and action recogonition and alerting using estimated pose. Open Pose library is used for the realtime 2D pose estima- tion and it consists of recognition of 18 body key points and joint locations and a final parsing to get skeleton. These are further used to extract robust motion features, then a Convolution Neural Network (CNN) is used to recognize the activities as sociated with these features. Different subjects from different camera angles are used to make the approach person-independent. The proposed method shows best result with promising performance, reaching an overall accuracy of 91.6% Thu, 01 Sep 2022 00:00:00 GMT http://210.212.227.212:8080/xmlui/handle/123456789/270 2022-09-01T00:00:00Z