Study Of Feature Extraction And Recognition Of Stored Product Pests

The stored grains, oils, seeds, furs, foods, including meats and aquatic products,tobaccos, medicines, furnitures, books, archives and the others are harmed by the storedproduct pests,which caused huge losses on our country’s agriculture and economic.Besides, the stored product pests are easily transported over long distances, which is athreat to human health. So it is very necessary for us to do the recognition andclassification of the stored product pests. It can facilitate intelligent monitoring andidentification of pests and prevent pest invasion.The traditional recognition of the stored product pests was mostly doing by expertswith experience. This works are intense and the recognition accuracy is low. But in thispaper the intelligent recognition method of the stored product pests can effectivelyimprove the recognition accuracy and efficiency, saving time and cost. The main contentsof this paper include the following aspects:(1) The biological analysis of stored product pests. By looking up a large amount ofrelevant information and analyzing morphological characteristics of stored product pests,the structural biology of pests is established, which laid a good foundation for thesubsequent feature extraction.(2) Image acquisition and pre-processing of stored product pests. Pest images arecollected by Shandong Inspection and Quarantine Bureau. These pest images should beprocessed by de-noising, segmentation and edge detection. In this paper, I use severaldifferent image pre-processing algorithms and compare the experimental results. We findthat the improved fourth-order partial differential equations method is good for pest imagede-noising, the result of threshold segmentation based on gray histogram and edgedetecting by contour tracking algorithm are better.(3) Feature extraction and selection of the stored product pests based on biologicalanalysis. These features include ten morphological features and seven HU invariantmoments. The species of the stored product pests are numerous and diverse and thefeatures are high-dimensional, so it is necessary to select these features by some featureselection algorithms. The feature selection algorithms include simulated annealing algorithm(SA), principal component analysis(PCA), Max-Relevance andMin-Redundancy (mRMR)algorithm. The processed features by the several featureselection algorithms are put into BP neural network classifier and support vector machineclassifier for recognition and classification.According to the classification results, we canbe seen that using the mRMR algorithm to select features is the best.(4) Analyzing the pros and cons of various classifiers. In this paper, BP neural networkand support vector machine are used to evaluate the relationship between image featurescombined with the recognition rate. Experimental results show that the BP neural networkrecognition rate is not high because the BP neural network is easy to fall into localminimum. Support vector machine uses four kinds of kernel function to model, theexperimental results show that the recognition rate is the highest by using the RBF kernelfunction.