基于图像检索的大豆食心虫虫害高光谱检测

为减轻虫害对大豆的影响，首先使用相应的高光谱仪器进行样本采集，样本分为4类：包括带有微小虫卵的，带有幼虫的，有啃食痕迹的和完全正常的大豆各20颗；然后提出了一种基于三维图像检索(3D-R-D，3D Resnet18 DCH)的大豆食心虫的高光谱检测方法。该方法从视频检索的应用中得到启发，考虑到视频不同帧之间和高光谱不同层之间存在类比关系，使用了在大规模视频检索数据集下训练而成的分类模型，将它作为预训练3D卷积模型进行训练。和已知的文献方法相同，使用公开的光谱数据集进行正式训练和微调，从而得到能进行特征提取的3D卷积网络，用图像检索来实现间接分类，通过利用样本之间的特征距离，实现在全新类别上的分类。为能适应任务，将模型最后的分类层变成了图像检索常用的hash层，从而得到了代表特征的二进制码。该方法不但完成了对不同情况下大豆种类的检测，还解决了训练时样本不足的问题。为探寻一种好的相似度匹配损失函数，对比了多种较新的方法，最后发现使用融入柯西分布的损失函数，实验效果最佳，最终模型的分类精度达86%±1.00%，和在大豆食心虫检测上最新的小样本方法对比，3D-R-D方法提高了3.5%左右的精度，表明该方法是有效的，它也为结合高光谱检测相关研究提供了一种全新思路。

Hyperspectral Detection of Soybean Heart-Eating Insect Pests Based on Image Retrieval

To reduce the impact of insect pests on soybeans, first use the corresponding hyperspectral instrument to collect samples. The samples are divided into 4 categories: those with tiny eggs, those with larvae, those with nibbling marks, and those with entirely normal soybeans. 20 Then, a hyperspectral detection method of soybean borer based on three-dimensional image retrieval (3D-RD, 3D Resnet18 DCH) was proposed. The application of video retrieval inspires this method. Considering the analogy relationship between different frames of the video and the different layers of the hyper spectrum, the classification model trained on the large-scale video retrieval data set is used as a predictive model. Train the 3D convolution model for training. Same as the known literature method, the public spectral data set is used for formal training and fine-tuning to obtain a 3D convolutional network that can perform feature extraction. Image retrieval is used to achieve indirect classification, and the feature distance between samples is used to achieve classification in new categories. In order to be able to adapt to the task, the final classification layer of the model is turned into a commonly used hash layer for image retrieval, thereby obtaining the binary code representing the feature. This method completes the detection of soybean types in different situations and solves the problem of insufficient samples during training. In order to explore a good similarity matching loss function, this article uses a variety of newer methods to explore and finally found that the Cauchy distribution is integrated into the loss function, which can be effectively applied in this experiment. Experiments have proved that the classification accuracy of the final model is 86%±1.00%. Compared with the latest small sample method in detecting soybean borer, the 3D-RD method improves the accuracy by about 3.5%, which shows that the method is effective. The method also provides a new way of thinking for hyperspectral research.