词袋特征PCA多子空间自适应融合的黄瓜病害识别

针对颜色纹理特征结合单一分类器的传统识别方法对于多种黄瓜病害的识别精度较低的问题,该文提出基于词袋特征PCA(principal component analysis)多子空间自适应融合的黄瓜病害识别方法。该方法首先对多种病害建立类别相关词袋模型,提取病害图像的高维词袋特征,然后用主成分分析法将病害高维特征降维到多个不同维数子空间,并在各子空间上分别训练BP(back propagation)神经网络;通过设置自适应阈值对待分类图像在各子空间上的分类得分进行融合得到识别结果。采集黄瓜角斑病、棒孢霉叶斑病、白粉病、霜霉病和炭疽病等5种常见病害部位图像共246幅,每类病害子图像中任意选择20幅作为训练集(共100幅),其余146幅作为测试集进行病害识别试验。结果表明在2个主成分子空间融合分类的平均准确率为90.38%,比传统颜色特征、纹理特征和颜色纹理混合特征识别率分别高6.97、26.15和13.02个百分点,且算法对不同类别病害的分类准确率更稳定,为温室黄瓜病害诊断提供了一种有效方法。

Bag of words feature multi-PCA subspace adaptive fusion for cucumber diseases identification

Abstract: Cucumber disease identification in greenhouse is of great significance to cucumber production. The traditional recognition method based on colour and texture features with single classifier has low accuracy in diagnosing many kinds of diseases. In view of this problem, a cucumber disease identification method based on Bag of Words (BoW) model and multi classifiers adaptive fusion on PCA-subspaces, BoW-mPCA, was proposed. First, for each category disease images, the SIFT points were extracted and clustered into a certain number of clusters, and the centres, called visual words (VW), were gathered to establish the category-related BoW (CR-BoW) model. Considering that the image data in the remote transmission can reduce the real-time of the disease recognition, and the disease location plays a decisive role in disease identification, so we cut the diseases location of the sample images by hand into sub-images of different size (minimum 58×68, maximum 562×487), which were collected as the dataset for CR-BoW model construction and experiments. To present the images based on the CR-BoW model, every SIFT point of each image was presented by the nearest VW in the VW space. Then each image was presented as a histogram of the VWs appearing on the image. To solve the problem of the high-dimension BoW-feature of disease images, PCA (Principal Component Analysis) method was adopted to reduce the feature dimensionality to series of subspaces of different size, and on each subspace a BP network was trained. During the identification process, a self-adaptive strategy for classifiers combination was built based on the idea that each single type of disease should be classified on one optimal PCA-subspace. The diseases image was classified by the BP net trained on the smallest PCA-subspace. If the distance of the highest two scores was higher than a given threshold, the image was classified into the category corresponding to the highest score. If not, the image was classified again by the BP net trained on the next larger PCA-subspace, and the scores were confused with the former ones and the highest two were compared to decide the final result. Repeating this process until the highest two scores were obviously discrepant or all the BP nets were used. The experiments were carried on a dataset containing five types of cucumber diseases, angular leaf spot, corynes pora, powder, downy and anthrax. The results showed that the average classification precision of five kinds of cucumber diseases in two PCA subspace fusion with preservation of 75%, and 85% energy, respectively, was 90.38%, higher than the single colour, texture and colour-texture mix feature by 6.97, 26.15, 13.02 percent respectively. Although for single disease of angular leaf Spot, corynes pora and anthrax, BoW-2PCA obtained precisions of 86.96% and 88.00%, slightly lower than colour features (89.78% and 93.20%), the precision on downy was 89.13% comparing to 64.20% from colour feature. This result indicated that BoW-2PCA could deal with the situation that the traditional methods can hardly handle. Moreover, the BOW-2PCA algorithm showed more stability in classification precision over all five types of diseases than the other three traditional methods. These results showed that the proposed CR-BoW model and BoW-mPCA algorithm was effective for diagnosing greenhouse cucumber disease.