Research On Plant Leaf Recognition Based On Two-stage Classification

Plants have a very wide distribution on the earth and are an integral part of nature.The classification of plants has a long history,but traditional plant identification can only be achieved through time-consuming and labor-intensive methods such as consulting books.With the rapid development of image processing and machine learning technologies,it has become possible to automatically recognize plants through computers.It is against this background that we have studied plant leaf identification methods in this paper.The main work of this article is as follows:1.A shape feature based on region area ratio is proposed.As an overall description of the plant leaf shape,the feature selects the circumscribed rectangle containing the leaf shape area as the target region,divides the target region into several equal parts,and then calculates the area ratio of the leaf shape in each region.The experiments prove that this feature has excellent description ability,and has the invariance of rotation,translation and scale transformation.2.A shape feature based on contour corner points is proposed.The occurrence frequency of the leaf contour corner points in each region is counted in the circumscribed circular of the leaf shape area to describe the leaf shape in detail.Experiments show that this feature can effectively describe the leaf shape details,and has the invariance of rotation,translation and scale transformation.3.A shape-based two-stage classification strategy is proposed.In the first stage of classification,plants are divided into several categories only based on the leaf shape.In the second stage of classification,the shape,texture,and color features are combined to accurately classify plants of each shape category.In order to enhance the description ability of plant leaves with similar shapes,this paper also proposes a method for plant leaf vein description.4.A series of experiments are performed on the Flavia dataset,Swedish dataset,ICL dataset and the LZU dataset.The algorithm proposed in this paper has achieved good performance in each dataset,and its classification accuracy rates have reached 99.3%,98.9%,91.5%,and 95.3%,respectively.The experiments verify the effectiveness of the features proposed in this paper,and prove that the feature combination and shape-based two-strategy classification strategy proposed in this paper is feasible.