论文部分内容阅读
日常生活中人们分拣辨别不同种类的苹果需要消耗大量的人力物力,为解决这一问题,提出了一种基于多角度多区域特征融合的苹果图像分类方法。首先,收集五类总共329个苹果,使用手机摄像头从上面、下面和三个不同侧面共五个角度采集每个苹果的图像,每个图像裁剪若干个(1~9)区域块;其次,每个区域块用颜色直方图向量来表示,多个区域块的直方图向量通过首尾相连进行融合,以此生成一个图像的表示;最后,将得到的329个样本数据用12种分类器进行分类比较。实验结果表明,当多角度多区域图像特征融合时,分类效果总是好于单角度单区域,而且越多越好;当使用五个角度的图像,每个图像裁剪9个区域时,PLS分类器的分类精度达到97.87%,好于深度学习。该方法操作简单,精度较高,算法复杂度为4n,n为图像裁剪区域块总数,可以推广成手机应用,并应用到更多水果和植物图像分类上。
In order to solve this problem, people in our daily life need a large amount of human and material resources to sort and distinguish different kinds of apples. To solve this problem, an apple image classification method based on multi-angle and multi-region feature fusion is proposed. Firstly, a total of 329 apples of five categories were collected, and the images of each apples were collected from the top, bottom and three different sides of the apples by using the camera of the mobile phone. Each image was cut into several (1-9) area blocks. Secondly, A region block is represented by a color histogram vector, and the histogram vectors of a plurality of region blocks are fused end to end to generate a representation of the image. Finally, the 329 sample data are compared and classified by 12 classifiers . The experimental results show that the classification results are always better than the single-angle single-area when the multi-angle and multi-area image features are fused, and the more the better, the best when using five angle images, and the nine areas are cut for each image. The classification accuracy of the instrument reached 97.87%, better than the depth of learning. The method is easy to operate and has high precision. The algorithm complexity is 4n. N is the total number of blocks in the image cropping area, which can be generalized into mobile applications and applied to more fruit and plant image classification.