基于卷积神经网络的番茄叶片病斑识别仿真

Simulation of Tomato Leaf Patch Recognition Based on Convolutional Neural Network

针对叶片病斑识别方法受噪声干扰影响较大,且Region Proposal和分类分作两块,导致mAP值较低的问题,在卷积神经网络基础上,研究一种新的番茄叶片病斑识别方法。对采集到原始番茄叶片病斑图像做灰度化处理,选择基于小波阈值的方法去噪,将去噪图像的直方图通过积分概率密度函数转化为概率密度为1(理想情况)的图像,利用卷积神经网络中的SSD算法将Region Proposal和分类统一在一起,提取与分类番茄叶片病斑特征,实现病害识别。结果表明,利用基于卷积神经网络的方法进行番茄叶片病斑识别,得到的mAP值为0.988,病斑识别识别性能更强,病害识别结果较为准确,为番茄病害治疗措施的选择提供了重要的理论依据。

In order to solve the problem of low map value caused by noise interference and the separation of region proposal and classification,a new method of tomato leaf lesion recognition based on convolution neural network was studied.At first,the original image of tomato leaf lesion was grayed and denoised based on wavelet threshold.Then the histogram of the denoised image was transformed into an image with probability density of 1(ideal situation)through integral probability density function.Moreover,SSD algorithm in convolutional neural network was used to unify region proposal and classification to extract and classify the features of tomato leaf lesions.Thus,the disease identification was completed.Simulation results prove that the map value of tomato leaf lesion recognition based on the proposed method was 0.988.In addition,this method has stronger recognition performance of tomato leaf lesion,and its disease identification result is more accurate.This method provides an important theoretical basis for the selection of treatment measures.