Disease identification in plants is very essential for preventing the loss in yield of the agricultural products. It is difficult to monitor the plant diseases manually at each stage, which requires more effort and time. The main objective of this work is to analyze and detect the diseases that affect the plants using image processing and machine learning techniques. The symptoms of plant diseases are identified and segmented using edge and color based image processing methods. Relevant features from the segmented diseased leaf portion are extracted and the type of disease is classified using multi-class Support Vector Machine. In this research work the plant diseases caused by different pathogens such as bacteria, fungus and virus are analyzed for early and periodical detection of plant diseases.
In this paper, Support Vector Machine (SVM) method has been employed to perform classification of brain tumor images into their types and grades. Mainly the focus is on four brain tumor types-Normal, Glioma, Meningioma, Metastasis and the four grades of Astrocytomas, which is a common type of Glioma. This work experimented the SVM classifier using different statistical feature set such as first order features, second order features and both together and the final analysis shows that for brain tumor types and grades classification, second order features alone achieve an accuracy of 85% and 78.26% respectively, whereas first order features alone achieve 65.517% and 62.31% respectively and when both are used together 84.48% and 68.1% accuracy is achieved respectively. The experiments are repeated for different SVM types, kernel types and gamma values of kernel type. Analysis on the misclassification is performed for each feature set using specificity and sensitivity measures. At the end of this work, we inferred that the SVM classification method is classifying the brain tumor types satisfactorily but comparatively lacks in tumor grade classification.
Image fusion is the process of combining relevant information from two or more images into a single fused image. The resulting image will be more informative than any of the input images. The fusion in medical images is necessary for efficient diseases diagnosis from multimodality, multidimensional and multiparameter type of images. This paper describes a multimodality medical image fusion system using different fusion techniques and the resultant is analysed with quantitative measures. Initially, the registered images from two different modalities such as CT (anatomical information) and MRI-T2, FLAIR (pathological information) are considered as input, since the diagnosis requires anatomical and pathological information. Then the fusion techniques namely Redundancy Discrete Wavelet Transform (RDWT) and Contourlet Transform are applied. Further the fused image is analyzed with five types of quantitative metrics such as Standard Deviation (SD), Entropy (EN), Overall Cross Entropy (OCE), Ratio of Spatial Frequency Error (RSFE), and Power Signal to Noise Ratio (PSNR) for performance evaluation. From the experimental results we observed that RDWT method provides better information (quality) using EN metric and the Contourlet Transform gives the difference in source to the fused image using OCE metric and also the fused image obtained from the proposed fusion techniques has more information than the source images are proved through all metrics.
Medicine is one of the major fields where the application of artificial intelligence primarily deals with construction of programs that perform diagnosis and make therapy recommendations. In digital mammography, data mining techniques are used to detect and characterize abnormalities in images and clinical reports. In the existing approaches, the mammogram image classification is done in either clinical data or statistical features of an image using neural networks and Support Vector Machine (SVM) classifier. This paper is proposed to evaluate the Application of Feature Extraction by means of combining the clinical and image features for clustering and classification in mammogram images. Initially, mammogram dataset is divided into training and test set. For the training and test sets, preprocessing techniques like noise removal and background removal are done to the images and Region of Interest (ROI) is identified. The statistical features are extracted from the ROI and the clinical data are obtained from the dataset. The feature set is clustered using k-means algorithm followed by SVM classification to classify the image as benign or malignant. The accuracy obtained from the proposed approach of clustering followed by classification is 86.11% which is higher than the direct classification approach where the accuracy is 80.0%. From the above results, the superiority of the proposed approach in terms of accuracy is justified.
In recent years, classification of human activities has gained more interest, this paper presents a novel analysis of a multivariate timeseries dataset namely PAMAP2 (physical activity monitoring) using different classification algorithms. The dataset consist of 18 day-to-day activities with 54 attributes that include sitting, standing, lying, ironing, vaccum-cleaning, walking, running, cycling, nordiac walking, playing soccer, rope jumping, car driving, ascending and descending stairs etc. of which first five activities are selected for analysis. The classification is carried out using Weka data mining tool on algorithms namely J48, Naives Bayes and Random Forest. The results proved an accuracy of 100%, 100%, 98% for training set and 19.2% 90.11% 93.2% for test set on J48, Random Forest and Naives Bayes algorithms respectively. Thereby, the overall analysis demonstrates that Random Forest algorithm outperforms better on PAMAP2 (physical activity monitoring) dataset and the activity, vaccum-cleaning results with highest correctly classified instances on Random Forest and Naives Bayes algorithms.