基于字典学习与SSD的不完整昆虫图像稻飞虱识别分类

为了解决图像采集过程中由于昆虫图像获取不完整而导致整体稻飞虱识别精度低、速度慢的问题，提出了一种基于字典学习和SSD的不完整稻飞虱图像分类方法。首先，使用自主研发的野外昆虫图像采集装置采集稻飞虱图像，构建小型图像集。然后，将采集的稻田昆虫图像进行阈值分割，得到单一稻田昆虫图像；对单一昆虫图像进行分块处理，得到带有背景信息和特征信息的混合子图像块集；使用子图像块作为字典原子来构建过完备字典，并对其进行初始化和优化更新；将更新后的过完备字典作为训练集输入SSD算法中进行训练，得到训练模型。最后，将采集的包含不完整稻田昆虫的图像在训练集模型上进行测试，并将测试结果与BPNN（Back propagation neural network)、SVM (Support vector machines）、稀疏表示等方法进行对比。试验结果表明，所提出的基于字典学习和SSD的稻飞虱识别与分类方法可以对不完整的昆虫图像进行准确快速的识别分类，其中，分类速度可达22f/s，识别精度可达89.3%，对稻飞虱的监督、预警和防治提供了有效的信息与技术支持。

Recognition and Classification of Rice Planthopper with Incomplete Image Information Based on Dictionary Learning and SSD

In order to solve the problem of low accuracy and slow speed of identification of rice planthopper caused by incomplete insect image in image acquisition process, a rice planthopper identification and classification method with incomplete insect images based on dictionary learning and single shot multibox detector (SSD) was proposed. Firstly, the field insect image acquisition device was used to acquire rice planthopper images and a small image set was built by these images. Then, single-inspect images were obtained by thresholding of the collected images of rice insects. Single-insect images were divided into blocks to obtain a mixed sub-image blocks with background information and feature information. The sub-image blocks were used as dictionary atoms to construct an over-complete dictionary, and this initial dictionary was optimized and updated immediately. The updated over-complete dictionary was trained as the training set of the SSD algorithm to obtain the training model. Finally, the collected incomplete insect images were tested on the obtained training models, and results were compared with back propagation neural network (BPNN), support vector machines (SVM) and sparse representation. Experimental results showed that the research on the identification and classification method with incomplete images based on dictionary learning and SSD can identify and classify rice planthopper accurately and quickly. Classification speed was 22f/s, the recognition accuracy was 89.3%. Hence, the method proposed can provide effective information and technical support for the supervision, early warning and control of rice planthoppers.