ABSTRACT. Identification and prevention of pest insect are essential requirements for proper crop cultivation. However, identifying pest insects can be a daunting and time consuming task because of the similarities of visual traits between different species. As a result, there are some necessities for a well performing automated system that can classify pest insects from image data. In this research, we propose a noble model that takes advantage of transfer learning and data augmentation to classify insect pest species from image data in the most accurate way. In the proposed model, three different Deep Neural Network (DNN) models were used for image classification: VGG19, Inception v3 and ResNet50. With appropriate data augmentation, Inception v3 achieved the best accuracy of 57.08% on a total of 102 insect species classification, beating the previous best result of 49.4% on the same dataset. Additionally, all the species were grouped based on the crops they consume. As Inception v3 was the best performing model across all the classes, it was also used to classify crop specific insect species. For 8 different crops, a range of approximate 48.2% to 88.1% accuracy was achieved from classification. Finally, all the results were analyzed, compared and discussed.

ABSTRACT. Guava disease has become a tremendous problem for the production of guava which has undeviating effect on the socioeconomic development of the farmers. This phenomenon leads to initiate an automated computer vision based guava disease detection system that may detect malicious guava and guide to early cure approaches, resulting reduction of relative economic loss. Considering the fact, in this paper we have proposed a convolutional neural network (CNN) based guava disease detection and curative suggestion providing system. We have collected the images of guava affected by anthracnose, fruit rot and fruit canker along with disease free guava from different districts of Bangladesh. In this paper, we have applied three CNN models and experimentally found that the third model has outperformed the other two with an accuracy of 95.61\%. For meticulous experimentation, performance metrics like precision, recall and F1 score is evaluated and found to yield great results.

ABSTRACT. Bangladesh is a land of agriculture, where people consumes rice as the main meal for three times a day. Rice is undoubtedly the most cultivated crop in Bangladesh. Like every other crops, rice also gets affected by a lots of diseases. These diseases differ from region to region and season to season. Although a number of implementation of different technology in agricultural field is increasing at an enormous rate, the farmers of our country still depends on the ancient techniques of disease identification. By keeping this very thing in our mind, we have conducted this research where we have tried to develop a model which can recognize rice diseases by deploying machine learning. We have worked with six main disease that is commonly seen in the paddy fields of Bangladesh. Authentic dataset of these six diseases were collected very carefully so that our model can render us the highest accuracy rate. BRRI(Bangladesh Rice Research Institute) has assisted us a lot in this matter. Three vastly popular pre-trained models of CNN such as Inception -v3, MobileNet-v1 and Resnet50, have been used to carry out this research. Necessary augmentation and scaling was done in the dataset before employing them. The research yields gratifying outcome. Hence, it proves that how effectively machine learning can collide with the agriculture. This research will pave machine learning techniques a path to enter in the agricultural sector of our country as well as help the young generation immensely who will enter into the agriculture in the future.

ABSTRACT. Rice is considered as the main food for about 140 million people in Bangladesh. Rice, as a food, does not only fulfill the protein or calorie intake of an average person, but also rice production plays a vital role in terms of rural employment and GDP of the country. However, the production of rice is hampered because of many diseases of rice leaves. The objective of this work is to develop a model which can predict those diseases so that farmers can take appropriate action. This work presents a CNN based model which provides 97.40% accurate results in predicting various diseases of rice leaves. Using a dataset of over 900 images of diseases and healthy leaves and following the technique of 10-fold cross-validation, the model was trained to identify 4 common rice diseases. This is the highest accuracy gained for only rice disease prediction to the best of our understanding with such a large dataset covering at least 4 diseases. The results of the simulation represent the feasibility and efficacy of the proposed model.

ABSTRACT. Market basket analysis finds out customers’ purchasing patterns by discovering important associations among the products which they place in their shopping baskets. It not only assists in decision making process but also increases sales in many business organizations. Apriori and FP Growth are the most common algorithms for mining frequent item sets. For both algorithms a predefined minimum support is needed to satisfy for identifying the frequent itemsets. But when the minimum support is low, a huge number of candidate sets will be generated which requires large computation. In this paper, an approach has been proposed to avoid this large computation by reducing the items of dataset with top selling products. Various percentages of top selling products like 30%, 40%, 50%, 55% have been taken and for both algorithms frequent item sets and association rules are generated. The results show that if top selling items are used, it is possible to get almost same frequent item sets and association rules comparing with that outputs which are derived by computing all the items within a short time. From time comparison it is also found that FP Growth algorithm takes smaller time than Apriori algorithm.

ABSTRACT. Disease that affects the cornea are the leading cause of corneal blindness. One of the major causes of corneal blindness is corneal ulcer. To diagnose corneal ulcer, biomedical machines like special microscope slit-lamp are mostly used. In recent years, digital image processing and machine learning techniques are widely used for automatic disease detection, diagnosis, and clinical decision-making procedure to achieve the optimum and most accurate results. In this paper, an automated method is proposed to detect corneal ulcer disease using convolutional neural network (CNN) from visual contents of facial image and also automatically segment the corneal ulcer area with high segmentation results. The eye parts are segmented from frontal facial image and classify corneal or non-corneal ulcer disease using CNN. For segmenting corneal ulcer area, the skin area is excluded from eye image and detect eye boundary using GrabCut method. We segment the iris and sclera region using Hough Gradient and Active contour techniques. We also apply several Morphological operations and Active contour techniques on iris image to detect exact ulcer location and to measure the ratio of ulcer size. Our method shows improved accuracy than other methods. The average accuracy rate of CNN is 98.78%.

ABSTRACT. Diabetes is one of the most common diseases that can affect anyone at any ages. These diseases attacked when the glucose level or sugar level is increased. Predicting diabetes is one of the most important things at this moment. There are some several techniques applied on Indian Pima Dataset. The dataset studying on woman Pima Indian population which had started in 1965. Most of the researcher trying to apply some complex techniques on dataset while many comprehensive research has many common technique missing. In our study, we have applied some very popular techniques such as Neural Network(NN), Support Vector Machine(SVM), Random Forest(RF), etc. We have applied those methods in several ways. Firstly, we have applied several algorithms in the original dataset. Then we used several preprocessing techniques to identify diabetes. Finally, we applied those techniques to compare and get the best accuracy. Neural Network was given the best accuracy(80.4%) than any other techniques.

ABSTRACT. Cancer has become one of the most life threatening disease over the past few decades. Especially on Bangladesh the number of people being affected by cancer is increasing in an agitating rate. Again cancer, diagnosed after a certain stage, inevitably leads towards death. To abate this vicious upheaval of cancer, awareness has no other alternative. Our research primarily focuses on detection of certain age group, according to the corresponding cancer diagnosis and relevant factors. In order to do so, we have implemented logistic regression, support vector machine and convolutional neural network on the original dataset. Afterwards, two feature selection methods (Feature Importance Ranking Method and Recursive Feature Elimination) have been applied on the dataset to extract out the most significant features. The three classifier comparison has been implied on both the feature selection methods. It is found that the classifier accuracy on the extracted features is significantly better in case of Recursive Feature Elimination rather than Feature Importance Ranking Method.

ABSTRACT. Diabetic retinopathy (DR) is the primary cause of vision loss among the grown-up people around the world. In four out of five cases having diabetes for a prolonged period of leads to DR. If detected early, more than 90% of the new DR cases can be prevented from turning into blindness through proper treatment. Despite having multiple treatment procedures available that are well-capable to deal with DR, it is the negligence and failure of early detection that costs most of the patients their precious eyesight. The recent developments in the field of Digital Image Processing (DIP) and Machine Learning (ML) have paved the way to use machines in this regard. The contemporary technologies allow us to develop devices capable of automatically detecting the health of a person’s eyes based on their retinal images. However, in practice, several factors hinder the quality of the captured images and impede the detection outcome. In this study, a novel early blind detection method has been proposed based on the color information extracted from retinal images using an ensemble learning algorithm. The method has been tested on a set of retinal images collected from various rural areas of South Asia, which resulted in a 91% average classification accuracy.

ABSTRACT. Software testing for a system is necessary not only to identify whether its attainment of required software qualities but also to identify whether it is defect-free or not. But there is not any standard automated testing cases generation framework in computer vision communities, especially in gesture recognition. This paper has proposed an automatic software testing cases generation framework to ensure the efficiency of the gesture recognition systems. Our goal is to build a standardized framework for testing the performances of existing gesture recognition systems. In our research, we have considered the ISO/IEC /IEEE 291129 -2013 standard for software testing process and ISO/IEC /IEEE 291129 -2015 for software testing techniques. In our proposed framework, we have considered five parameters such as rotation, contrast, scaling, background, and noise which are used to generate test cases based on existing renowned gesture recognition systems. We have selected five gesture recognition systems as experimental purposes. The test process calculates accuracy using our generated test cases and compares with the existing systems results. Finally, a comparative analysis is given to improve the efficiency of the system.

ABSTRACT. In this paper, a 2x2 multiband patch antenna array is presented and it is designed and simulated by using CST Microwave Studio for the bands 24GHz-27GHz, 31- 34GHz, 42GHz-48GHz, 50GHz-52GHz, 65GHz 68GHz, 82GHz-85GHz. The design of patch antennas is very efficient and widely used in wireless communication due to their lower cost of fabrication, light weight and can operate a microwave frequency but it offers low efficiency, low gain etc. Future upcoming 5G wireless communication is needed of high gain, good protection from path loss because of their millimeter wavelength, broadband performances of antennas. Similarly, the multiband antennas are highly desired to use a single antenna instead of multiple antennas for different bands. The antenna provides maximum and minimum gains of 15.63dB and 9.23dB at operating frequency 26.82 GHz and 34.60 GHz. The radiation efficiency is achieved 96% and above for all operating frequencies, total bandwidth of 11.84 GHz.The overall size of antenna is 20.66mmX22.56mmX0.67mm make it normed to work better by fulfilling the requirements for 5G communication.

ABSTRACT. The computational technique for solving continuous algebraic Riccati equations governed from very large dimensional power system with sophisticated ingredients requires highly expensive time dealings and invade by the infeasible rate of convergence. Aim of the work is mainly focused on acquiring the optimal control for the large-scale power system model and stabilize the corresponding system through the Riccati based feedback stabilization. To achieve the desired goal, a projection based nested iterative algorithm is proposed by means of rational Krylov subspace method. The proposed algorithm will allow the structure preservation simulations and can be efficiently applied to the perturbed systems without preconditions.

ABSTRACT. Voltage controlled oscillator (VCO) is one of the key components required for the wireless communication systems. In this paper a low power, low phase noise 3-stage differential ring VCO is designed for passive Radio Frequency Identification (RFID) transponders in microwave frequency range that is compatible with Zigbee, IEEE 802.11 b/g and Bluetooth protocols. 90 nm CMOS technology is adopted and the simulations of the designed VCO is performed in Cadence virtuoso environment. Simulation results shows that it has wide tuning range from 1.04 GHz to 4.21 GHz (120.7 %) and the power consumption is 2.27 mW at 2.45 GHz. The oscillator has a phase noise from -92.01 dBc/Hz to -86.90 dBc/Hz at 1 MHz offset frequency. At 2.45 GHz oscillation frequency the figure-of-merit (FOM) of has a value of – 153.35 dBc/Hz. Additionally, to assess the VCO in diversified environments process corner analysis, temperature sweeping, Monte Carlo analysis and stability analysis has been performed. Although the proposed VCO is designed for passive RFID tags, the architecture is suitable for many other wireless applications for microwave frequency.

ABSTRACT. Electromyography (EMG) is a way of measuring the bioelectric activity of muscles. EMG is usually performed to detect abnormalities in the nerves or muscles of a target area. The recent developments in the field of Machine Learning allow us to use EMG signals to teach machines the complex properties of human movements. At present, machines are capable of detecting numerous human activities and distinguishing among them solely based on the EMG signals produced by those activities. However, success in accomplishing this task primarily depends on the learning technique used by the machine to analyze EMG signals, and even the latest algorithms do not result in flawless classification. In this study, a novel classification method has been described employing a multichannel Convolutional Neural Network (CNN) that interprets surface EMG signals by the properties they exhibit in the power domain. The proposed method has been tested on a well-established EMG dataset, and the result yields very high classification accuracy. This learning model will help researchers to develop prosthetic arms that capable of detecting various hand gestures, and mimicking them afterwards.

ABSTRACT. This paper presents a deep neural network to design inset fed rectangular shaped microstrip patch antenna. To design the shape of inset fed rectangular patch antenna, a multilayer perceptron based deep neural network has been proposed based on artificial neural network to predict the dimension values. The proposed model converged closely with 2% mean squared error over the network output parameters of antenna. Hence the proposed model can be used to design rectangular shaped microstrip patch antenna at any given frequency.