基于卷积神经网络的农业害虫图像匹配点识别方法

A Method Recogniting Matching Points on Agricultural Pest Image Based on Convolutional Neural Networks

当前病虫害图像智能识别过程提取的局部特征主要突出图像细节,对图像中光照等外界环境干扰较为敏感,匹配精度低,提出一种基于卷积神经网络的农业害虫图像匹配点识别方法。将害虫标本与活体图片作为训练集与测试集,建立卷积神经网络,在卷积层中,上个层级的特征经卷积核处理后,利用激励函数即可获取新一层特征。池化层选择max poling进行处理,学习过程采用梯度反向传播形式。在分类层中,针对目标函数,通过Momentum动量法进行处理。采用建立的数据集对卷积神经网络特征视觉词袋模型进行训练,依据虫害图像特征,通过对处于同一叶子节点上的匹配点进行相似度计算,完成对匹配点的识别。实验结果表明,在视角发生改变的情况下,本文方法、SURF方法、BRIEF方法的匹配比率没有很大的差异,而在季节改变与光照改变的情况下,本文方法的匹配比率明显优于SURF方法、BRIEF方法。

At present,the local features extracted in the process of intelligent recognition of pest image mainly highlight the details of the image,which is sensitive to the interference of external environment such as light in the image and has low matching accuracy.A method of agricultural pest image matching point recognition based on convolution neural network is proposed.In the convolution layer,the features of the previous level are processed by convolution kernel,and the new level features can be obtained by using the excitation function.In the pool layer,Max poling is selected for processing,and gradient back propagation is used in the learning process.In the classification layer,the objective function is processed by momentum method.Based on the characteristics of pest image,the matching points on the same leaf node are calculated by similarity,and the matching points are identified.The experimental results show that the matching ratio of this method,surf method and brief method is not very different when the angle of view changes,but the matching ratio of this method is obviously better than surf method and brief method when the seasons change and the light changes.