Detecting Of Chestnut Classification Based On Multi-Source Information Fusion Technique

Chestnut is an important economic crop in China, however, at present, the overall quality of Chinese chestnut is seriously affected by unqualified chestnuts with mildew, insects or other defects. Currently, removal of defective chestnuts mainly depends on manual sorting, which is time-consuming, labor-intensive and inefficient. Especially, due to chestnut's complex shape, it is difficult to discern the internal and external defects chestnut caused by mildew and insects. Therefore, to develop a fast, accurate, non-destructive method for identification of defects in chestnuts is of important scientific significance and great practical value.Experimental chestnuts were sampled from Jingshan area in Hubei province. Near-infrared spectroscopy and machine vision technology were studied in detection and classification of chestnut. By using near infrared spectroscopy technology, defective and qualified chestnuts were analyzed, and five spectral data pretreatment were compared for their effects on classification results, and the discrimination model was studied; a chestnut image acquisition system was built, and Chestnut grading model was studied based on machine vision technology. To improve the detection precision, data fusion technology of near-infrared spectroscopy and machine vision for Chestnut grading was proposed and nondestructive methods of Chestnut classification was established based on multi-source information fusion technology. The main conclusions are as follows:1. Classification and detection model of chestnut was established based on near infrared spectroscopy technique. Using principal component analysis (PCA), spectral feature parameters of chestnut were extracted. Spectral feature parameters of chestnut samples in training set, which contain 240, were extracted as the inputs of pattern recognition. Grading models were built based on near infrared spectroscopy and BP neural network. The validation set which contains 80 chestnut samples was adopted to verify the model. Experimental results showed that the discriminating rate was 96.25% in training set, and 86.25% in prediction set.2. A chestnut image acquisition system was established based on computer vision technology. Programmers for image processing and analysis were developed on MATLAB. Median filter was the optimal noosing method. Chestnut image and background were segmented by edge detection and thresholding method for grey image segmentation. Feature parameters of Chestnut image, such as color, texture and defect, can characterize the difference of defective and qualified chestnut in color, texture and other features.Using principal component analysis (PCA), image feature parameters of chestnut were extracted. Image feature parameters of chestnut samples in training set composed of 240 samples were taken as the inputs of pattern recognition. Grading models were built by machine vision and BP neural network, the validation set which contains 80 chestnut samples was adopted to verify the model. Experimental results showed that the discriminating rate was 96.67% in training set, and 83.75% in prediction set, respectively.4. Multi-information fusion technology was used to establish non-destructive testing and classification model of Chestnut by data fusion of near infrared spectroscopy and machine vision. Spectral and image feature parameters of chestnuts were fused at feature level. Further, Chestnut classification model based on multi-source information fusion tecchnology was built by employing BP neural network and least squares support vector machine fusion method, respectively.Spectral and image feature parameters of chestnut were extracted by using principal component analysis (PCA). Feature parameters of chestnut in training set, which was composed of 240 samples, were extracted as the inputs of pattern recognition model. The model is of BP neural network structure with three layers of 9-15-2.80 samples in the test set were used for validation. The results indicated that correct recognition rates of chestnut grading model based on multisensor information technology and BP neural networks were 97.92% and 90% for training set and test set, respectively.Spectral and image feature parameters of chestnut were extracted by using principal component analysis (PCA). Feature parameters of chestnut in training set, which was composed of 240 samples, were extracted as the inputs of pattern recognition model. The model is of BP neural network structure with three layers of 9-15-2.80 samples in the test set were used for validation. The results indicated that correct recognition rates of chestnut grading model based on multisensor information technology and BP neural networks were 97.92% and 90% for training set and test set, respectively.