基于稀疏编码和SCGBPNN的鳞翅目昆虫图像识别  

Using sparse coding and SCG BPNN to recognize images of lepidopteran insects

作　　者：竺乐庆[1] 张真[2] 

机构地区：[1]浙江工商大学计算机与信息工程学院,杭州310018 [2]中国林业科学研究院森林生态与保护研究所,国家林业局森林保护重点实验室,北京100091

出　　处：《昆虫学报》2013年第11期1335-1341,共7页Acta Entomologica Sinica

基　　金：浙江省自然科学基金项目(LY12F02048);浙江省教育厅资助项目(Y201119748)

摘　　要：【目的】为了给林业、农业或植物检疫等行业人员提供一种方便快捷的昆虫种类识别方法,本文提出了一种新颖的鳞翅目昆虫图像自动识别方法。【方法】首先通过预处理对采集的昆虫标本图像去除背景,分割出双翅,并对翅图像的位置进行校正。然后把校正后的翅面分割成多个超像素,用每个超像素的l,a,b颜色及x,y坐标平均值作为其特征数据。接下来用稀疏编码(SC)算法训练码本、生成编码并汇集成特征向量训练量化共轭梯度反向传播神经网络(SCG BPNN),并用得到的BPNN进行分类识别。【结果】该方法对包含576个样本的昆虫图像的数据库进行了测试,取得了高于99%的识别正确率,并有理想的时间性能、鲁棒性及稳定性。【结论】实验结果证明了本文方法在识别鳞翅目昆虫图像上的有效性。[ Aim ] In order to find a convenient way to recognize insect species for those worked in agriculture, forestry, plant quarantine etc., we developed a novel method to recognize images of lepidopteran insects. [ Methods ] Firstly, the background of captured specimen image is removed and then the wings are cut out and calibrated in the preproeessing period. Then the calibrated wing is segmented into a number of super pixels, and mean values of l, a and b in color space and x and y in Cartesian coordinate system are kept as feature data. Following that, the sparse coding （SC） algorithm is used to train the codebook, generate the sparse codes that are pooled into a feature vector to train the SCG （Scaled Conjugate Gradient ） Back Propagation Neural Network （BPNN）. Finally the resulting BPNN is used to classify and recognize unknown insects. [ Results ] The proposed method was tested in a database with 576 images with the best recognition rate over 99%, and the system also demonstrated ideal time performance, good robustieity and stability. [ Conclusion ] The experimental results proved the efficiency of the proposed method in recognizing images of lepidopteran insects.

关 键 词：昆虫 鳞翅目 图像识别 超像素分割 稀疏编码 量化共轭梯度法 反向传播神经网 

分 类 号：Q969.42[生物学—昆虫学] TP391.41[自动化与计算机技术—计算机应用技术]