Welcome to visit Scientia Silvae Sinicae,Today is

Scientia Silvae Sinicae ›› 2018, Vol. 54 ›› Issue (5): 78-86.doi: 10.11707/j.1001-7488.20180509

Previous Articles     Next Articles

A Method for Forestry Business Images Classification Based on Auto-Learning Features

Li Yingjie1, Zhang Guangqun2, Wang Hangjun1   

  1. 1. Jiyang College, Zhejiang A & F University Zhuji 311800;
    2. School of Information Engineering, Zhejiang A & F University Lin'an 311300
  • Received:2016-09-05 Revised:2017-02-22 Online:2018-05-25 Published:2018-06-05

Abstract: [Objective] With the popularization of modern electronic and communication equipment, large numbers of forestry business images can be gotten in time and be accumulated in forestry administration. Automatic classification of forestry business images can achieve effective supervision of forest resources because it is the core of decision support, rapid reaction and cooperation among all forestry management departments, including forestry law enforcement agencies and law enforcement teams.[Method] A new method for forestry business image classification, which was based on auto-learning local features was proposed. Three layers of networks are constructed in the method. They are used to get the local features of images, to form global semantic by combine these local features, and to model classifiers. Firstly, patches are extracted from images and their features are learned automatically using linear sparse auto-encoder. The features are local. Then, the images are convoluted with the local features to get the global activation feature maps. The maps are then pooled and combined to form feature vectors. Finally, softmax is used to model and classify on the vectors.[Result] Four categories are established for forestry business images. They are animal death, forest fire, logging and forest pest. 355 images are gathered for the four categories with different numbers of images in different categories. They are cropped to the same size and constitute the experimental dataset. Leave-one-out cross validation is done on the dataset and we get the 80% classification accuracy. And, it can be seen that different color features, color variation features, gradient features and other features are extracted automatically from image patches by linear sparse auto-encoder automatically. We also can see that different features are activated in different feature maps which are gained by convoluting images with the weight matrices. Also, the active features are scattered in feature maps. An image will be identified correctly, if the features, which are the same or similar with those in other images of the same category, are included in this image. This illustrate that our method performs well.[Conclusion] The issue of forestry business images classification is a kind of scene classification problem. The images resemblance in the same category is weak. To classify them with traditional "feature extraction and classification modeling" method,the appropriate classification features are difficult to find. By comparison, our method learns the local features automatically and it is more general. Also, only three network layers are included in this method, and there is no feedback learning crossing all these three layers. So, it is more efficient than the popular methods which are based on deep convolutional neural networks. Higher accuracy can be expected while the accumulating images are more and comprehensive.

Key words: forestry business images, linear sparse auto-encoder, convolution, pooling, softmax

CLC Number: