A Novel Approach for Generic Plant Disease Classification Puneet Mugabasav, Shreyas M. Salotagi, Akash Kulkarni, Kaushik Mallibhat 1st International Conference on Advances in Computer Science Electrical Electronics and Communication Technologies Ce2ct 2025, 2025 With advances in the globalization of the world market, several new varieties of crop cultivation have popped up. The advances have led to habitat changes, and new varieties of diseases and plants have surfaced to challenge the farmer community. Several advancements have been made to assist today's agriculture. One among them is the utilization of learning-based models to identify plant diseases. However, the literature review shows that studies were carried out to identify specific diseases in particular plants. The exacting models failed to achieve cross-identification of the diseases irrespective of the type of plant. Thus, the proposed study focusses on novel classifier intending towards generic plant disease classification across various crops. The study was trained and tested for mildew, spot, and rot diseases. Through the training and testing for seven plants and blind validation across six other plants, an accuracy of 96.43% in testing and 63% accuracy in validation was obtained.
Problem Based Learning – An Effective way of teaching DSP course Rohit Kalyani, N. S R, G. H. M., R. Shet, Nalini Iyer, Satish Chikkamath, Akash Kulkarni, S. Budihal Journal of Engineering Education Transformations, 2024 The Digital Signal Processing (DSP) course is fundamental to many signal processing and communication applications. Since it is a mathematical intensive, faculty often face challenge in teaching and keeping students engaged for longer time. Students find it challenging to understand the core concepts and practical applications. The present work investigates Problem-Based Learning (PBL) approach to enhance the teaching of DSP by integrating real-world problems and active learning strategies. This new approach was offered to third-year students of the university, combining PBL with traditional lectures. The curriculum is redesigned according to constructive alignment. It included core DSP concepts as modules, PBL activities for each module and their assessment. The PBL was implemented in different phases focusing on orientation, problem identification, signal analysis, and system design. The new experience resulted in improved student engagement, concept understanding, and problem-solving skills in DSP, with positive feedback from students. The future plan is to focus on addressing the challenges faced during the initial implementation and to refine the PBL approach for the next batch of students. Keywords— Digital signal processing (DSP); PBL; Real world problems, Reflection; Self-directed learning; Traditional teaching;
Analysis on Utilization of Learning Model In-place of Traditional Demodulation Block Sahana Kolur, Bhargav Hegde, Sumati Javalagi, Akash Kulkarni Proceedings Icnews 2024 2nd International Conference on Networking Embedded and Wireless Systems Wireless Technology Building A Digital World, 2024 The need for adaptable modules in the communication system is ever-high in this computationally advancing time. Great efforts are being made to advance the modulation technique to achieve a higher rate of error-free data transfer. Advancements in using learning models in the modulation-demodulation blocks are limited. Here, our study to achieve the demodulation block with a sampler and CNN model is analysed, and results are presented. The model works almost similarly to the expected theoretical values of the error probability. The model aligns with its function compared to the theoretical model for the SNR between -5 and 0dB. However, if trained and built to achieve an extensive range of SNRs and different modulation schemes simultaneously, it functions as an adaptive demodulation block all in one learning module block. This module is observed to demodulate and detect the received bits better than a traditional demodulation block.
Analysis of Estimation Accuracy between Kalman and LSTM Algorithm S P Vijay, J PrajwalGouda, Nishchit Mogali, Akash Kulkarni, Kaushik Mallibhat 2024 5th International Conference for Emerging Technology Incet 2024, 2024 Our work intends to compare the Kalman and LSTM algorithms on their ability to estimate the next state. In today's data world, a large base of information is available, and many studies are going on to establish the accurate predictor of future states, say, the stock market, earthquake, weather forecasting, etc., that play important roles in today's life. Kalman filter is a main estimating algorithm used in automotive rockets and satellite landings. However, the same confidence is not given to the learning models. This work intends to study the same system dependence of Kalman and develop the same confidence in the RNN model. The results show that RNN is 9.76% more accurate than Kalman for rapidly varying data. Presenting the main advantage of RNN to handle non-linear variations in the data.
Lane Detection for Autonomous vehicles using Image Transformation Techniques Poonam M Shettar, Abhijeet Jadhav, Divya Karoshi, Aishwarya G, Akash Kulkarni 2024 5th International Conference for Emerging Technology Incet 2024, 2024 Lane detection is critical in autonomous driving and advanced driver assistance systems (ADAS), furnishing vital information for vehicle navigation and safety. The study introduces lane detection methodology leveraging established image processing techniques within the Image transformation frameworks, including the study of kernels. This approach accurately detects lane markings in real-time images or video streams with the help of Gaussian blur and Canny edge detection. The system's evaluation primarily focuses on structured roads under standard conditions, demonstrating its efficacy in such environments. However, the potential for enhancing vehicle autonomy and safety across varied driving scenarios remains prominent. Leveraging Image transformation capabilities and the insights gained from kernel studies, this research advances computer vision applications in the automotive sector, facilitating the evolution of more intelligent and adaptable driving systems. Furthermore, the study introduces a method for assessing the accuracy of the detected lanes by calculating the intersection over union (IOU).
Digitalising Handwritten Hindi Documents Using Learning Models Kumar K K, Kruthik Gupta B, Mahesh R Kallur, Akash Kulkarni, Kaushik Mallibhat 2024 IEEE 9th International Conference for Convergence in Technology I2ct 2024, 2024 The character classification for handwritten documents has many challenges since each individual write with different font, has a different style of writing and writes with fontsize. This unconventional style poses a challenge to traditional techniques like Optical Character Recognition (OCR) and the solutions exists at the level of recognizing the character and a word. Through this project, author extends it for entire document digitalization thus helping to preserve the important documents written in Hindi. The authors used Convolution Neural Network models to address the challenge and were able to achieve 99.19% accuracy in classifying the characters.
Galaxy Type Classification With Sub-categorization Using Deep Learning Framework Aryan Kamat, Nitish Kumar Pathak, Aditya Phatak, Akash Kulkarni, Kaushik Mallibhat 2024 IEEE 9th International Conference for Convergence in Technology I2ct 2024, 2024 This paper presents a robust approach to galaxy categorization utilizing machine learning and image processing techniques, emphasizing Galaxy Zoo and Galaxy10 datasets. The MobileNetV2 architecture is used for feature extraction on a training dataset of 4100 labeled photos. At the same time, image pre-processing procedures ensure dataset standardization. After 20 epochs, the model reaches an overall accuracy of 91%, indicating its usefulness in categorizing galaxies into Elliptical, Spiral, and Irregular types, as well as detailed subclassifications such as E0, E3, E7 for elliptical galaxies and Normal Spiral, Barred Spiral for spiral galaxies. The subclassification models use image processing to extract morphological traits, demonstrating the possibility for automated study of astronomical information and contributing to advances in astrophysics research.
Adaptive Headlight Control using Q-learning Reinforcement Algorithm Rahul Hegde, Tarun Divatagi, Shreyas Mutnal, Akash Kulkarni, Kaushik Mallibhat 2024 5th International Conference for Emerging Technology Incet 2024, 2024 In this paper, we introduce an adaptive headlight control system using reinforcement learning to overcome challenges faced in the automobile industry. The project aims to ease visibility issues using an adaptive system, considering the varying climatic conditions during day and night driving. The system uses Q-learning to automatically change the brightness levels to the surrounding environment and classify intermediate brightness levels that were not observed during training, also efficiently classifying recognized “light” and “dark” intensity ranges. This quantitative evaluation of the model's performance demonstrates how flexible and capable it is at making decisions in new situations. The concept shows great promise in improving safety and maximizing energy efficiency in car lighting systems overall.
