Plant Image Sets Classification Using Manifold And Convex Hull Modeling

Plant taxonomy is the science that finds, identifies, describes, classifies, and names plants. With the rapid development of pattern recognition technology, the plant image recognition technology has played an important role in plant taxonomy and brought great convenience to scientific research and production.In traditional plant image classification techniques, after extracting various image features, classification and recognition are often based on the single sample classifier. The same class species also express different state in different environments, temperature, time, etc. Even if the big difference between leaves or flowers in just one plant, therefore, the classification based on single image has significant limitations. Nowadays, for each person it is easy to collect multiple plant images in digital age, each class of plant images make up an image set. These image sets are captured in different scenes and times, covering different conditions and variations. Comprehensive comparisons and extensive experiments show plant image sets base recognition has great performance.We first extract a variety of image feature after pretreatment of plant images, and then model the image sets by two methods. First we model the each set of images as a manifold, and in the frame of manifold-manifold distance, proposed plant images classification based on manifold-manifold distance. Specifically, we apply a clustering procedure in order to express a manifold by a collection of local linear models. Then the distance is measured between local models which come from different manifolds that constructed above. Finally, the problem is transformed to integrate the distance between pairs of subspace. Experiment shows the method has a great performance.The second method is image set modeled by convex hull. We propose adaptive multi affine hull classification method. Solve the problem of unreasonable non-convex data modeling, and use only a small number of boundary information, susceptible to noise data in single affine hull modeling. First, modeled the test set as multi affine hull using spectral clustering, then using adaptive reference clustering to filter out the noise images, finally compared the similarity of local convex hulls.