CN104280349A - Method for identifying hollowness of white radishes based on hyperspectral image - Google Patents

Abstract

The invention relates to a method for identifying the bran heart of white radish based on a hyperspectral image, which belongs to the non-destructive detection technology in the storage and processing industry of agricultural products. Through the hyperspectral imager, the transmission hyperspectral image of white radish during storage is obtained, the difference in spectral response between normal white radish and bran heart white radish is analyzed, and the spectral value in the wavelength range of 400-1000nm is extracted as the input value of the neural network. Find out whether the white radish has a bran heart. This method can realize the accurate identification of the bran heart of white radish, replace manual destructive detection, effectively prevent unqualified products from flowing to the market, improve the edible and processing utilization rate of white radish, promote the development of deep processing industry of radish, and provide a basis for the application of hyperspectral technology in the field of agricultural products. learn from.

Description

A method for identification of white radish bran heart based on hyperspectral images

technical field

The invention relates to a hyperspectral image technology method for detecting bran core during post-harvest storage of white radish, and belongs to the technical field of non-destructive detection of agricultural product storage and processing.

Background technique

Radish bran heart, also known as hollow, is a natural phenomenon in the growth of radish, and it can occur during both the growth period and the storage period. There are many reasons for the bran heart of radish. Water imbalance, unsuitable fertilizer conditions, light and temperature will all lead to bran heart of radish. The bran core process will reduce nutrients such as starch and sugar, and affect its processing, storage and edible properties. The traditional detection method of bran heart in radish is to use artificial sensory detection, which is not only time-consuming and laborious, but also has low precision, which is difficult to meet the needs of large-scale industrial automatic grading. Therefore, it is of great significance to establish a non-destructive and reliable method to detect the bran core of radish, to detect and grade radish, to improve the market value of radish and to develop the deep processing industry of radish.

In recent years, hyperspectral image detection technology has been widely recognized as a non-invasive and rapid method for analyzing and evaluating the quality and safety of various foods. Hyperspectral images can detect the physical and morphological characteristics of food, as well as internal chemical and molecular information, so as to analyze and evaluate the quality and safety of food. This technology has been well applied in the food industry at home and abroad, such as Jianwei Qin et al [Qin J, Burks T F.Development of a two-band spectral imaging system for real-time citrus canker detection[J].Journal of Food Engineering, 2012, 1(108): 87-93.] Based on the characteristic bands of hyperspectral image screening, a commercial fruit classifier was developed with a speed of 5 pieces/second and an overall classification accuracy of 95.3%. Ana Herrero-Langreo et al [Herrero-Langreo A, Lunadei L, et al.MultispectralVision for Monitoring Peach Ripeness[J]. Food science, 2011, 2(76): 178-187.] using hyperspectral image technology to evaluate peach ripeness Degree, to facilitate the determination of the best picking time. Piotr Baranowski et al [Baranowski P, et al. Detection of early bruises in apples using hyperspectral data and thermal imaging [J]. Journal of Food Engineering, 2012, 3(110): 345-355.] using hyperspectral images to detect apple hardness and Soluble solids were assessed. Hyperspectral imaging technology has also been applied to the detection of surface defects in apples, cherries and citrus fruits, internal defects in cucumbers, etc. In recent years, domestic use of hyperspectral image technology to detect the quality of agricultural products has also developed rapidly, such as Huang Qianwen et al. .2013, 29(1): 272-277.] Detection of slight damage on apple surface, Gao Hailong et al. Chinese Journal of Agricultural Engineering. 2013, 29(15): 279-285.] Detection of potato black heart disease, Li Jiangbo et al. Research [J]. Spectroscopy and Spectral Analysis. 2012, 32(1): 142-146. Using hyperspectral fluorescence to detect early rotten navel oranges, Tian Youwen et al [Tian Youwen, Li Tianlai, Zhang Lin, et al. Diagnosis of greenhouses using hyperspectral image technology Cucumber disease methods [J]. Agricultural Engineering Journal. 2010 (5): 202-206.] Cucumber disease detection and other aspects have achieved good results. However, the technology of non-destructive detection of bran core inside radish has not been reported at home and abroad. It is necessary to carry out research on non-destructive detection of bran core inside radish using hyperspectral image technology.

Contents of the invention

technical problem

In view of the above-mentioned technical development status, the purpose of the present invention is to develop a fast and non-destructive method for hyperspectral image detection to meet the urgent needs of the radish deep-processing industry in view of the problem that the existing technology cannot realize the non-destructive identification of the bran heart during storage and sales. By using hyperspectral imaging technology, the spectral information difference between normal white radish and white radish with bran heart was analyzed, the characteristic parameters of the response were extracted, and the identification model of bran heart white radish was constructed.

Technical solutions

1. A method for identification of white radish bran heart based on hyperspectral image, its device is characterized in that:

1) System composition, including hyperspectral imaging unit, sample holder, motorized platform, line light source, light box, computer and image acquisition software, the whole device is placed in a closed black box. Among them, the hyperspectral imaging unit consists of a camera (Imperx, ICL-B1620, with a wavelength range of 400-1000nm and a spectral resolution of 2.8nm), a spectrometer (Specim, ImSpector, V10E) and a variable focal length lens. The light box is a 150W The halogen tungsten lamp is transmitted by a linear fiber optic tube. The computer model is CPU E5800, 3.2GHz, memory 2G, graphics card 256M GeForce GT240; image acquisition software is Spectral Image software independently developed;

2) Transmission acquisition unit, the distance between the lens and the white radish sample is 20cm, the sample is placed close to the line light source, the light source intensity is 90W, the acquisition exposure time is 70ms, the acquisition speed is 2.5mm/s, and the image resolution is 804×440;

Its detection steps are:

1) Take a sample of white radish to be tested with uniform size and no mechanical damage, clean the surface and dry it, and place it in the hyperspectral image detection system to obtain a hyperspectral image;

2) Use the following formula to correct the obtained image to obtain the corrected hyperspectral image:

$\mathrm{<span\; class="notranslate">\; Rc</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; D.</span>}}{\mathrm{<span\; class="notranslate">\; W</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; D.</span>}}$

Among them: Rc is the corrected hyperspectral transmission image, _R0 is the original hyperspectral transmission image, W is the standard white calibration plate with a reflectivity of 99.99%, placed directly above the light source, scanning the transmission white board to obtain a full white calibration image , D is to cover the lens with the lens cover, and collect a completely black calibration image;

3) Select the region of interest with 25,000 pixels in the middle of the white radish area in the image, extract the spectral mean value of all pixels in the region in the range of 400-1000nm, and use it as the input value of the constructed neural network model, and output The result is whether the white radish is bran-hearted, and the parameters of the neural network model constructed are: the number of hidden layers is 1, the number of nodes in the hidden layer is 13, the activation function of the hidden layer is hyperbolic tangent; the number of output layers is 2, that is, the number of qualified samples and Bran core samples, the output layer activation function is Softmax.

Beneficial effect

The present invention utilizes the monitoring of the response signal of the hyperspectral image instrument to accurately distinguish whether the bran heart inside the white radish is through the hyperspectral characteristics of the white radish without destroying the integrity of the white radish, which can standardize product quality and improve The market value of radishes and the reduction of consumers' concerns about the possibility of buying radishes with bran heart are of profound significance to the radish deep processing industry. Compared with traditional destructive testing, it not only saves time, but also avoids unnecessary waste. The technology and method are novel, and the research results can not only be used for rapid analysis and detection in the laboratory, but also can be used for the identification of white radish bran heart in industrial automatic production through the development of online detection equipment and portable instruments, as well as for the internal quality of other agricultural products. The detection provides a useful reference.

4. Description of drawings

Figure 1: Device for identification of white radish bran heart by hyperspectral transmission system

Figure 2: Device for identification of white radish bran heart by hyperspectral reflectance system

Figure 3: Device for identification of white radish bran heart by hyperspectral semi-transmission system

Figure 4: Raw average spectrum of white radish in transmission mode

Figure 5: Raw average spectrum of white radish in reflectance mode

Figure 6: Raw average spectrum of white radish in semi-reflective mode

5. Specific implementation

A method for identifying the bran heart of white radish based on hyperspectral images, the specific implementation is as follows:

1. Test material

The white radish variety is the 301 radish selected by the Jiangsu Academy of Agricultural Sciences. It was cultivated on May 20, 2014. During the planting process, according to the cause of the radish bran heart, some radishes were specially treated to cause the bran heart. On July 10, 2014 Harvested at No. 1, 120 samples with no mechanical damage, no pests and diseases, and the same size were selected, including 60 treated and untreated samples, cleaned and dried naturally for testing.

2. Hyperspectral image acquisition system

The hyperspectral imaging system is mainly composed of cameras, imaging spectrometers, CCD cameras, light sources, a set of mechanical conveying devices, and computers, and is produced by Taiwan Isuzu Corporation. The spectral effective band range of the imaging spectrometer is 400-1000nm, a total of 440 bands, the spectral resolution is 2.8nm, and it is equipped with a focal length variable lens. The light source is a 150W halogen tungsten lamp with a total of 10 levels, adjustable, and transmitted by optical fiber to line lights. In order to avoid the influence of external light on the spectrum collection, the detection device is placed in a dark box as a whole, and the background is black without reflection. The test uses three detection modes of transmission, reflection and semi-transmission to acquire hyperspectral image signals of white radish. The hardware configuration of the three acquisition modes is the same, and the difference is the relative position of the light source and the sample and the parameter settings.

The hyperspectral image acquisition system based on the transmission mode is shown in Figure 1. The sample and light source are fixed on the conveyor belt, and a line light source is located directly below the sample. The light passes through the sample and is absorbed by the spectrometer, and converted into data and sent to the computer. The relevant parameters are set as exposure time 70ms, conveyor belt speed 2.5mm/s, light source intensity 90W, light source close to the sample, camera lens 20cm away from the sample, fixed sample to prevent rolling, and start data collection.

The hyperspectral image acquisition system based on reflection mode is shown in Figure 2. Two line light sources are respectively fixed directly above the sample, and the sample is fixed on the conveyor belt. The beams emitted by the line light sources intersect, and the intersection point of the beams is the center of the fruit. Collect images to collect spectral information, convert it into data and send it to the computer. The relevant parameters are exposure time 3ms, conveyor belt speed 3mm/s, light source intensity 45W, angle of line light source 45°, camera lens distance 25cm from sample, fixed sample to collect data.

The hyperspectral image acquisition system based on the semi-transmission mode is shown in Figure 3. Two line light sources are located on both sides of the sample, the sample is fixed on the conveyor belt, and the light beam emitted by the line light source is directly injected into the sample. The spectral information diffusely reflected by the sample is converted into data and transmitted to the computer. The relevant parameters are set as exposure time 45ms, conveyor belt speed 3mm/s, light source intensity 75W, light source placed horizontally close to the center of the sample, camera lens 25cm away from the sample, fixed light source and sample to start collecting data.

3. Hyperspectral image acquisition and correction

In order to eliminate the noise in the data collection process, under the same conditions as the white radish sample collection, a white standard calibration plate (99.99% reflectance) was scanned to obtain a completely white calibration image, and a completely black calibration image was obtained after covering the lens cover. The acquired absolute image is converted into a relative image by the formula, the formula is:

$\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; B</span>}}{\mathrm{<span\; class="notranslate">\; W</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; B</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

In formula (1): R is the converted relative image, I is the collected absolute image, B is the all-black calibration image, and W is the all-white calibration image.

During data processing, the region of interest analysis method was used to collect and convert the hyperspectral image of each sample, and the average spectrum of the region of interest (ROI region) with 25,000 pixels in the middle position was selected as the spectral value of the sample, and combined with the full spectrum Partial least squares analysis (PLS-DA), support vector machine (SVM), and artificial neural network (ANN) were used to model and distinguish bran heart radish.

4. Mathematical modeling method

Using three methods of PLS-DA, SVM, and ANN for modeling, comparing the spectral information obtained by the three detection modes under each model for the classification ability of white radish bran heart, comparing the classification accuracy rate, and judging the best detection mode and prediction model .

5. Raw spectrum analysis in different acquisition modes

The white radish with bran heart will change its tissue structure and chemical composition, which will affect the optical properties such as light transmission and absorption, which is quite different from that of radish without bran heart. Therefore, through the spectrum (transmission, reflection, The difference in semi-transmission) is expected to be used to determine whether the radish is bran heart. Collect the region of interest of the white radish hyperspectral image, calculate its average value, and obtain the spectral value of the white radish in the 400-1000nm band interval. Figures 4, 5, and 6 are the comparison charts of the average spectra of normal radish and bran heart radish in transmission, reflection, and semi-transmission modes, respectively.

Comparing the spectral curves obtained by the three detection modes of the hyperspectral image, it can be seen that the spectra of normal radish and bran heart radish are quite different, showing the possibility of hyperspectral image detection of radish bran heart, which can be used to distinguish normal white radish from bran heart Heart white radish. Moreover, since the white radish has a white surface and strong spectral reflection, the transmission and semi-transmission spectra can fully interact with the interior of the radish, which can better show the difference between normal radish and bran-heart radish.

6. Identification of white radish bran heart under different models of full-band spectrum

Using PLS-DA to distinguish bran heart radish from normal radish, the results are shown in Table 1. It can be seen that the PLS-DA model based on the transmission mode has an overall discrimination accuracy of 85.0% and 90.0% for the normal radish in the modeling set and prediction set, and 97.5% and 90.0% for the bran heart radish. , so the overall recognition accuracy of the transmission mode under the PLS-DA model is relatively high, and the effect of distinguishing bran core is good. In the reflection detection mode, the PLS-DA model has an overall recognition accuracy of 97.5% and 90.0% for normal radishes in the modeling set and prediction set, respectively, and 65.0% and 75.0% for bran heart radishes, which is normal in this detection mode. The recognition accuracy of radish and bran-heart radish is unstable. Only 3 qualified samples are identified as bran-heart samples, while 19 bran-heart samples are misjudged. Therefore, the reflection mode cannot judge whether the radish is good under the PLS-DA model. Bran heart. In the semi-transmission mode, the PLS-DA has an overall recognition accuracy of 72.5% and 75.0% for normal radishes in the modeling set and prediction set, and 97.5% and 70.0% for bran-heart radish samples respectively. Under the -DA model, the recognition rate of normal radish and radish with bran heart is low, and the effect of distinguishing bran heart is not good.

Table 1 Prediction results of white radish bran heart by PLS-DA model

Using SVM to distinguish bran heart radish from normal radish, the prediction results are shown in Table 2. It can be seen from the table that the SVM model based on the transmission mode has an overall discrimination accuracy rate of 100.0% and 95.0% for the normal radish in the modeling set and prediction set, and 100.0% and 85.0% for the bran heart radish. , under the SVM model, the overall recognition accuracy of the transmission mode is relatively high, and the effect of distinguishing bran core is good. In the reflection detection mode, the SVM model has an overall recognition accuracy of 87.5% and 75.0% for normal radishes in the modeling set and prediction set, respectively, and 95.0% and 90.0% for bran heart radishes. The correct rate is lower than that of bran-heart radish, and the overall ability of the reflection mode to identify radish bran-heart under the SVM model is average. In the semi-transmission mode, the overall recognition accuracy of SVM for normal radishes in the modeling set and prediction set was 80.0% and 75.0%, respectively, and the recognition accuracy rates for bran-heart radish samples were 92.5% and 70.0%, respectively. The semi-transmission mode under the SVM model The recognition accuracy of normal radish and bran-heart radish is not stable, and it is impossible to distinguish radish bran-heart.

Table 2 Prediction results of white radish bran heart by SVM model

Using ANN to distinguish bran heart radish from normal radish, the prediction results are shown in Table 3. It can be seen that the overall discrimination accuracy of the ANN model based on the transmission mode is 100% and 94.4% for the normal radish in the modeling set and the prediction set, and 97.7% and 94.1% for the bran heart radish. The overall recognition accuracy of the transmission mode under the ANN model is relatively high, and the effect of distinguishing bran core is good. In the reflection detection mode, the ANN model has an overall recognition accuracy of 88.1% and 100.0% for normal radishes in the modeling set and prediction set, respectively, and 76.9% and 85.7% for bran heart radishes. Bran heart radish recognition rate is average. In the semi-transmission mode, the overall recognition accuracy rates of ANN for normal radishes in the modeling set and prediction set were 84.8% and 71.4%, respectively, and the recognition accuracy rates for bran heart radish samples were 81.8% and 93.8%, respectively. The semi-transmission mode under the ANN model The recognition rate of normal radish and bran heart radish is low, and the effect of distinguishing bran heart is not good.

Table 3 Prediction results of white radish bran heart by ANN model

7. Comparison of the recognition effect of three prediction models on the heart of radish bran

From the above, it can be found that different detection modes and prediction models have differences in the detection of radish bran heart. Using the PLS-DA model to identify the bran heart of radish, the overall accuracy of the modeling set and prediction set in the transmission mode reached 91.3% and 90.0%, but the recognition accuracy in the reflection mode was unstable, and the recognition accuracy in the semi-transmission mode was relatively low. Low; using the SVM model to distinguish the bran heart of radish, the overall accuracy rate of the modeling set and prediction set in the transmission mode reached 100.0% and 90.0%, and the accuracy rate has been improved to a certain extent, but the recognition accuracy rate in the semi-transmission mode is low and not good. Stable; using the ANN model to distinguish the bran heart of radish, the overall accuracy of the modeling set and prediction set in the transmission mode reached 98.8% and 94.3%, and the recognition accuracy rate was high and the stability was improved compared with the previous two models.

Therefore, on the whole, among the three detection modes, the hyperspectral transmission detection mode has the highest accuracy in detecting white radish bran heart, and the overall accuracy of PLS-DA, SVM, and ANN models for the recognition of bran heart is the highest, obviously. Superior to reflective and semi-transmissive modes. Among the three prediction models, SVM and ANN have higher accuracy rates, but considering the accuracy and stability, the ANN prediction model has the best effect in distinguishing white radish bran.

Claims (1)

1. A method for identification of white radish bran heart based on hyperspectral images, the device is characterized in that, 1) The system consists of a hyperspectral imaging unit consisting of a camera, a spectrometer and a variable focal length lens, a sample holder, a motorized platform, a line light source, a light box, a computer and image acquisition software. The entire device is placed in a closed black box. Among them, the camera is Imperx, ICL-B1620, the wavelength range is 400-1000nm, the spectral resolution is 2.8nm, the spectrometer is SpecimV10E; the light box is a 150W halogen tungsten lamp, and the transmission is completed by a linear optical fiber guide; the computer model is CPU E5800, 3.2GHz, memory 2G, graphics card 256M GeForce GT240; image acquisition software is self-developed Spectral Image software; light source is transmission mode, in which the distance between the lens and the white radish sample is 20cm, the sample is placed close to the line light source, and the light source intensity is 90W, acquisition exposure time 70ms, acquisition speed 2.5mm/s, image resolution 804×440; 2) The detection steps are: 1. Take the white radish sample to be tested without mechanical damage, the surface is clean and free of debris, and placed in the hyperspectral image detection system as claimed in claim 1, to obtain hyperspectral images; ② Use the following formula to correct the obtained image to obtain the corrected hyperspectral image: $\mathrm{<span\; class="notranslate">\; Rc</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; D.</span>}}{\mathrm{<span\; class="notranslate">\; W</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; D.</span>}}$ Among them, Rc is the corrected hyperspectral transmission image, _R0 is the original hyperspectral transmission image, W is a standard white calibration plate with a reflectivity of 99.99%, placed directly above the light source, and scanning the transmission white board to obtain a full white calibration image , D is to cover the lens with the lens cover, and collect a completely black calibration image; ③ Select the region of interest with 25,000 pixels in the middle of the white radish area in the image, and extract the spectral mean value of all pixels in the region within the 400-1000nm band range, as the input value of the constructed neural network model, and the output result is Whether the white radish is bran heart, the parameters of the neural network model constructed are: the number of hidden layers is 1, the number of nodes in the hidden layer is 13, the activation function of the hidden layer is hyperbolic tangent; the number of output layers is 2, that is, the number of qualified samples and bran heart Sample, the output layer activation function is Softmax.