Study On Detection Methods Of Internal And External Defects Of Potato Based On Different Hyperspectral Imaging

With the implementation of the plan of potato staple food, the potato industries has been rapid development, the potato quality problem has become a hot issue. Potato quality will directly affect the development of the potato processing industry, the market and other economic benefits. Therefore, carrying out a rapid method for detecting potato quality lossless has important scientific significance.In this paper, the potato of Shandong Shuyingyihao was cited for the study. The semi-transmission hyperspectral imaging technology, reflective hyperspectral imaging technology and multi-so,urce information fusion technology was used to detect the normal, germination, light green, holes and black heart under any posture with one detection model.At the same time,a single index detection method for light green potato was build. The results are as follows:(1) Determined the potato internal and external quality detection method based on reflection hyperspectral imaging technology. The 489 potatoes(normal 122, germination103, lightly green 103, holes 103, black heart 58) were selected as study samples and collected the potato samples reflection hyperspectral images.In the image-dimensional, sample image was extracted RGB, HSV and Lab color space information. Compared the multi-classification model effect based on the 3 kinds of manifold learning dimensionality reduction methods, that were Isometric Mapping(Isomap), Maximum Variance Unfolding(MVU) and Laplacian Eigenmaps(LE).Ultimately, LE was determined as the optimal reflection image information dimensionality reduction methods. The deep belief network(DBN) was used to establish the potato classification detection model based on reflection hyperspectral image information. In the testing set,the model mixed recognition rate was 80.98 %, the single sample recognition rates of germination, green leather, holes, black heart were 95%,88.57%, 69.70%, 77.14% and 65%.In the spectral dimension, 4 kinds of methods that were Autoscale, standard normal variable correction(SNV), multiplicative scatter correction(MSC), and smoothing was used to preprocess the reflectance spectra. Compared the effect of the multi-classification model.Autoscale was determined as the optimal reflectance spectra preprocessing method. Further, the manifold learning dimensionality reduction methods locality preserving projection(LPP), local tangent space alignment(LTSA), local linear coordination(LLC) were used to reduce the dimension of reflectance spectrum thereafter Autoscale pretreated. The DBN model was established based on LPP, LTSA and LLC.Compared the recognition rate of the model based on different dimension reduction methods. Finally LTSA was the optimal dimension reduction method of reflectance spectra information. In the test set, the model mixed recognition rate was 87.73 %, the single sample recognition rates of germination, green leather, holes, black heart were95%, 85.71%, 81.82%, 91.43% and 80%.(2) Determined the potato internal and external quality detection method based on semi-transmission hyperspectral imaging technology. 5 kinds of potato including normal,germination, lightly green, holes, black heart were selected as study samples. Collecting the potato samples semi-transmission hyperspectral image.In the image dimension, extracting the image color information from RGB, HSV,Lab image space and comparing the DBN model based on the manifold learning dimensionality reduction methods, that are Isometric Mapping(Isomap), Maximum Variance Unfolding(MVU) and Laplacian Eigenmaps(LE). Finally the Isometric was determined as the optimal semi-transmission image information dimensionality reduction methods. In the test set, the model mixed recognition rate was 84.05%, the single sample recognition rates of germination, lightly green leather, holes, black heart were 95%,85.71%, 75.76%, 88.57% and 65%.In the spectral dimension, LPP, LTSA and LLC was respectively used to reduce the dimension of semi-transmissive spectral information, and build the DBN model thereafter MSC processed. Ultimately LTSA was determined the best dimension reduction method of the semi-transmission spectral information. In the test set, the model mixed recognition rate was 92.02%, the single sample recognition rates of germination,lightly green leather, holes, black heart were 97.5%, 88.57%, 84.85%, 94.29% and 95%.(3) Determined the potato internal and external quality testing methods based on hyperspectral images and spectral information fusion. Test with normal, germination,lightly green, holes and black heart five categories potato for the study. Combined the features, that were semi-transmission image feature with Isomap dimensions reduced, the reflection image feature with LE dimensions reduced, the reflection and semi-transmission spectral information with LTSA dimensions reduced into a new integration feature. Then the fusion models based on semi-transmission images,reflection images, semi-transmission spectra, reflection spectra, semi-transmission images and spectra, reflection images and spectra were build. Compared and analysed the6 fusion models, determined that the model based on semi-transmission images and spectra was optimal on detecting potato quality. The model mixed recognition rate was(4) 98.16%, respectively, normal, germination, lightly green, holes and black heart recognition rates were 100%, 90.91%, 100%, 100% and 100%, of the test set.(5) The mathod of detecting lightly green potato based on different hyperspectral imaging modalities was determined. In light of the difficulties in detecting the slightly green potatoes that were placed randomly, this paper take 225 potatoes as the study samples, including 122 normal potatoes, 103 lightly green potatoes. Collecting the reflection and semi-transmission hyperspectral image, and extracting color information,spectral information from the images, respectively. The Isomap, MVU and LE were utilized to reduce the image information dimension. The LPP, LTSA and LLC were used to reduce the spectral information dimension. The DBN was utilized to build the lightly green potato recognition model based on different hyperspectral image, spectra and multi-source information. The fusion model based reflection spectra information and semi-transmission spectra information,that wre reduced by LTSA was the best, after comparison. This model test sets recognition rate reached 100%. Non-distractive detecting of the slightly green potatoes can be realized with this fusion model.