Research On Rice Pest Identification And Counting Based On Image

The rice is one of the most important food crops in China, and the improvement of riceyield and quality is an important goal of rice production. However, in recent years, there's aaggravating tendency in rice pests outbreak year by year. Therefore, the control of rice pests inrice production and national economic development holds a very important status. Currently, theroutine method of rice pests monitoring relies mainly on the black light lamps for trapping thepests. The trapped pests are taken back, recognized and counted by technicists the following day.This method is of high labor intensity, low efficiency, poor objectivity and non-real time. Ithasn't satified the current demand of rice pests monitoring, because of serious pests outbreak. Toreduce the labor intensity of plant protection technicists and improve pests prediction ability, theidentification and counting of rice light-trap pests are studied on basis of machine vision, imageprocessing and pattern recognition technology in this paper. The main results are as follows:(1) An image acquisition system of indoor rice light-trap pests was developed. Firstly, pestsneeded to be preliminary screened through mesh screens with different size meshes. Twodigital cameras could capture both sides of all pests. A software was developed based onVisual C++6.0for automatically collecting pest images. The system was a good platform forautomatic recognition of the rice light-trap pests.(2) A new method was proposed for background segmentation based on gray difference. Theimages of mixed pests and its background were acquired in the same environment andcamera parameters settings. The results showed that this method could not only effectivelyavoid the edge fracture and empty holes phenomenon, but also possess the better robustness.(3) In the feature extraction, the HSV space was considered. The inhomogeneous quantificationhistogram was used to extract color features and it improved the recognition robustness.Some morphological features with rotating, translation and scale invariance were extracted toensure the recognition stability. The gray level co-occurrence matrix was used to extract textfeatures. It reduced the3/4calculation and dropped the complexity of feature extractionthrough compressing the grayscale of pest image.(4) The principal component analysis method was used to reduce the pests feature matrix'sdimension. The variance contribution of each main component was computed. According tothe rule which the accumulative contribution of the principle couldn't be less than85%, thefirst six principal components were selected as the features of pest identification. It couldeliminate the relevance of different variables and reduce the burden of computer memory.(5)7-fold cross validation was used to divide the dataset. The dataset was evenly divided to7 groups, a group was as test set and the remaining six groups were as training sets for trainingthe classifier. The result showed that this method could solve over-learning and owe-learningphenomenon.(6) In this paper, the multi-objective mixed rice pests were simultaneously recognized which isdifferent from individual pest recognition. Two methods, template matching and supportvector machine were used to recognize four rice pest species. The results showed that themulti-template matching method could reduce the influence of incomplete pests and postureon pest identification compared to single-template matching method. The pest recognitionrate of Support Vector Machine (SVM) method was14.4%higher than that of multi-templatematching. The SVM classifier should be selected to identify rice pests, because it has moreadvantages in the small sample, non-linear and high dimensional recognition, and thisalgorithm could avoid non-convergence and random of output result.(7) Rice pest counting was easy if these pests were identified accurately. The counting resultwould provide data sources for pests forecasting.Automatic recognition and counting of rice light-trap pests hasn't been researched before.The results of this paper presented a new idea for rice light-trap pest recognition. However, thereare many technical challenges, such as the segmentation of touching pests, the recognition ofincomplete pests and overlapping pests, and rejection of non-objective pests which neededdifferent experts to solve.