Linear Bezier Curve Geometrical Feature Descriptor for Image Recognition

Abstract

Background:Image retrieval has a significant role in present and upcoming usage for different image processing applications where images within a desired range of similarity are retrieved for a query image. Representation of image feature, accuracy of feature selection, optimal storage size of feature vector and efficient methods for obtaining features plays a vital role in Image retrieval, where features are represented based on the content of an image such as color, texture or shape. In this work an optimal feature vector based on control points of a Bezier curve is proposed which is computation and storage efficient. Aim: To develop an effective and storage, computation efficient framework model for retrieval and classification of plant leaves.Objective:The primary objective of this work is developing a new algorithm for control point extraction based on the global monitoring of edge region. This observation will bring a minimization in false feature extraction. Further, computing a sub clustering feature value in finer and details component to enhance the classification performance. Finally, developing a new search mechanism using inter and intra mapping of feature value in selecting optimal feature values in the estimation process.Methods:The work starts with the pre-processing stage that outputs the boundary coordinates of shape present in the input image. Gray scale input image is first converted into binary image using binarization then, the curvature coding is applied to extract the boundary of the leaf image. Gaussian Smoothening is then applied to the extracted boundary to remove the noise and false feature reduction. Further interpolation method is used to extract the control points of the boundary. From the extracted control points the Bezier curve points are estimated and then Fast Fourier Transform (FFT) is applied on the curve points to get the feature vector. Finally, the K-NN classifier is used to classify and retrieve the leaf images.Results:The performance of proposed approach is compared with the existing state-of-the-artmethods (Contour and Curve based) using the evaluation parameters viz. accuracy, sensitivity, specificity, recall rate, and processing time. Proposed method has high accuracy with acceptable specificity and sensitivity. Other methods fall short in comparison to proposed method. In case of sensitivity and specificity Contour method out performs proposed method. But in case accuracy and specificity proposed method outperforms the state-of-the-art methods.Conclusion:This work proposed a linear coding of Bezier curve control point computation for image retrieval. This approach minimizes the processing overhead and search delay by reducing feature vectors using a threshold-based selection approach. The proposed approach has an advantage of distortion suppression and dominant feature extraction simultaneously, minimizing the effort of additional filtration process. The accuracy of retrieval for the developed approach is observed to be improved as compared to the tangential Bezier curve method and conventional edge and contour-based coding. The approach signifies an advantage in low resource overhead in computing shape feature.