Description And Application Of Gaussian Convolution Angle Of Target Image

It is a basic problem in the research field of computer vision to effectively describe target images for subsequent target recognition tasks.Describing leaf image patterns for plant species identification and extracting butterfly image features for butterfly species identification are important applications in target image description.The plant leaf image mode and butterfly image mode generally have small inter-class differences and large intra-class differences,the shape distortion and self-occlusion of the target,as well as the change of illumination in image acquisition,the complex structure inside the target image,etc.,all bring great difficulties to the target recognition task.In this paper,a invariant descriptor called Gaussian convolution angle is proposed and applied to the recognition of plant leaf images and butterfly images.The Gaussian convolution angle proposed in this paper consists of a Gaussian convolution curve angle describing the shape feature,a Gaussian convolution patch angle describing the internal patch structure of the target,and a Gaussian convolution appearance angle describing the appearance feature of the target.Gaussian convolution curve angle is the included angle of left and right convolution vectors generated by convolution of left and right neighborhood chord vectors derived from contour points with Gaussian function respectively;Gaussian convolution patch angle is the included angle between left and right convolution patch vectors generated by projecting patch functions onto the chord vectors of left and right neighborhoods to generate patch vectors,and then convolution with Gaussian functions;Gaussian convolution appearance angle is the angle between the left and right convolution appearance vectors generated by projecting the gray scale function to the chord vectors in the left and right neighborhoods,calculating the gray scale statistical characteristics of each chord vector,generating appearance vectors,and then convolution with Gaussian function.The Gaussian convolution angle proposed in this paper has the following advantages:(1)It satisfies the invariance of translation,rotation,scaling,mirror image transformation and illumination change(theoretical proof is given in this paper);(2)It has the characteristic of multi-scale description.By changing the scale parameters of Gaussian function,multi-scale Gaussian convolution angles can be generated;(3)The target image is comprehensively described,and the method effectively describes theshape,complex internal structure and appearance characteristics of the target image.In order to verify the effectiveness of Gaussian convolution angle description method,this paper tests the performance of this method with the published plant leaf image databases: CVIP100 leaf image database and MEW2012 leaf image database.And its robustness is tested on the public Kimia data set.Experimental results show that this method has higher recognition rate and is more robust than the existing popular plant leaf image recognition methods.Different kinds of butterfly images have high shape similarity,but have different patch structures and appearance modes.In this paper,Gaussian convolution patch angle and appearance angle are applied to the identification of butterfly images,and the identification performance is verified by the published butterfly image test set.