Zou X.,Nanjing Agricultural University |
Zou X.,Jiangsu Province Engineering Laboratory of Modern Facility Agriculture Technology and Equipment |
Ding W.,Nanjing Agricultural University |
Ding W.,Jiangsu Province Engineering Laboratory of Modern Facility Agriculture Technology and Equipment |
And 4 more authors.
Nongye Gongcheng Xuebao/Transactions of the Chinese Society of Agricultural Engineering | Year: 2014
The rice planthopper images acquired by remote real-time recognition system usually have poor quality, and hence it is impossible to classify rice planthoppers using the color features of rice planthopper images. This study proposed to extract texture features of images based on gray level co-occurrence matrix (GLCM) and used the texture features to classify rice planthoppers. A H-shape mobile photographing device designed by us was used to obtain color images of rice planthoppers. The color images were grayed by formula, and then the background of images was removed using Otsu image segmentation method to generate binary images followed by calculation through the binary image coordinates. The GLCM was improved to extract texture features of images without background. Specifically, the center of gravity was determined by coordinates of the images and considered as the center to construct GLCM. The images of the rice planthopper were copied into the sub images with 160 pixels×160 pixels based on the center. Using multiple annular routes, the features of rice planthopper gray images were extracted including energy, entropy, moment of inertia and correlation. In the training and testing experiment of the extracted features, back propagation (BP) nerve network and optimized BP nerve network based on parametric selection -improved particle swarm optimization algorithm were individually used to train and classify the rice planthopper, and the training time and identification rate of each method were compared. A total of 300 Sogatella, Laodelphax and Nilaparvata lugens with 100 samples for each type of rice planthopper was trained. The training time using the optimized BP nerve network based on improved particle swarm optimization algorithm was only 0.5683 seconds, which was far less than that (29.5772 seconds) using BP neural network. Based on the BP neural network, the identification rate reached 80% for Sogatella, 90% for Laodelphax, and 95% for Nilaparvata lugens. Based on the improved particle swarm optimization algorithm-optimized BP nerve network, the identification rate reached 90% for Sogatella, 95% for Laodelphax, and 100% for Nilaparvata lugens. Therefore, the identification rate of the optimized BP neural network based on parametric selection-improved particle swarm optimization algorithm was higher than that of BP neural network. Furthermore, the shorter training time using the optimized BP neural network based on parametric selection-improved particle swarm optimization algorithm than using the BP neural network suggested that the former could better meet the requirement of real time optimization. Source