The use of soft computing to classification of some weeds based on video processing

作者：Sabzi Sajad; Abbaspour Gilandeh Yousef; Javadikia Hossein

来源：Applied Soft Computing, 2017, 56: 107-123.

DOI：10.1016/j.asoc.2017.03.006

• Classification
• Video processing
• Feature extraction
• Potato
• Machine vision
• Imperialist competitive
• Artificial neural networks

For long, the efficient use of herbicides has engaged the minds of many researchers. The traditional method where the entire farm is uniformly sprayed entails two major drawbacks: 1) Human health threat due to excessive consumption of herbicides; and 2) Environmental pollution. Therefore, in this study, a video processing system was proposed to detect Agria potato plants as well as Chenopodium album L., Secalecereale L, and Polygoniumaviculare L. The algorithm of such system performs spraying operation depending on the type and number of weeds. Videos were takenon4 ha of Agria potato fields in Kermanshah, Iran (longitude: 7.03 degrees E; latitude: 4.22 degrees N). The vision system of the proposed machine has two main subsystems. The first subsystem is responsible for making videos of the potato farms under controlled conditions. The second subsystem classifies potato plants along with the weeds of Chenopodium album L., Secale cereal L., and Polygoniumaviculare L. using the hybrid artificial neural network-imperialist competitive algorithm (ANN-ICA) classifier. After the recording and pre-processing of the objects, 25 features were extracted from each object. In this study, the decision tree algorithm was employed so as to select the most distinctive features. The eight selected features were area, length, width, perimeter, elongation, compactness, length to perimeter ratio, and image ratio. In order to classify the potato plants and different weeds based on these 8 features, the data were divided into two groups: Training and validation data with 2163 samples (15% for validation), and the testing data with 1015 samples. To compare the performance of the hybrid propagation neural network-imperialist competitive algorithm classifier, the meta-heuristic classifier of learning vector quantization (LVQ) and classic classifier of K-nearest neighbor (K-NN) were used. Sensitivity, specificity, and accuracy were considered in analyzing the performance of the confusion matrix. Except for the potato class, these values were above 91% for the hybrid ANN-ICA method. As far as LVQ method, these values were zero for the in Secale cereal L, and Polygoniumaviculare L. classes while they were less than 88% for the potato class. Finally, the results showed the superiority of the hybrid ANN-ICA method over the LVQ and K-NN methods, further corroborating the ability of the present method to detect the real-time of potato plants and three types of weeds.

• 出版日期2017-7