| It is a new technology to introduce the neural network and the spectrum analyzing technology into biology engineering and medical diagnosis in recent years. It is a new subject for the research of biologic interior through the detection and analysis of the biologic surface's spectrum. The research makes the technology of the optics, electronics, computer, and biological medicine together, so it has a bright future in food engineering, biological medicine and material engineering. In this paper, on the basis of transmission theory, the neural network theory and chromatic theory, several problems had been discussed as followed: The chromatic measuring system with fiber-optic probe had been designed; The adjusting methods had been designed for the attenuation introduced by the fiber-optic probe; The reflected spectrum of the apple samples had been measured and been classified with chromatic calculation; The neural network had been founded to analyze the kinds of apple samples; The fresh levels had been classified through the chromatic calculation and the neural network. So many achievements had been gained as followed: 一,The influence of the fiber-optic probe to the spectrum was analyzed systematically, and three methods had been put forward: Ratio adjusting method, weight ratio adjusting method and Fourier Transform spectrum of frequency adjusting method, and their performances had been compared with each other. A spectral analysis system with a fiber-optic probe is designed to enhance interference resistance ability of the chromatic system. This system includes light source, light-collecting system, light-splitting system, photo electronic transferring system and data-processing system. Light source is a bromide-tungsten light (color temperature is 2856K) which was standardized. A fiber-optic probe is used to collect and transfer light energy which is the light-collecting system. The probe can collect weak energy very well by reflective method. So this system has very close configuration, a stable performance and suitable for the measurement of micro-area. A hologram grating is controlled by a stepping motor, works as light-splitting system. Photoelectric multiplier tube, the photo electronic transferring system, is capable of measuring weak signals. The wavelength error is below 0.4nm and which proves that the system answers for the CIE's demands. Fiber-optic probe, which is taken as light-collecting system, it brought the distortion to the system. The characteristics of optic fiber attenuation in the range of visible light (380nm~780nm) is that the attenuation in red and purple light is much heavier than green light and no manifest moving. This paper put forward three adjusting methods as followed: 1. Ratio adjusting method is to measure five spectrum lines without fiber-optic probe. They are normalized and the average line can be gained. Similarly we also can get the average spectrum through the fiber-optic probe. So we can get the ratio adjusting function. 2. Weight ratio adjusting method is similar to ratio adjusting method, the difference between them is that every curve must be multiplied by a weight number which is between 0 and 1, in the former, and the weight reflected the contribution every curve to the average line, so the error had been lowered. 3. Fourier Transform spectrum of frequency adjusting method is to complete the compensation on the spectrum attenuation in generalized spectrum of frequency. It is to transform the output electric current with and without fiber-optic probe into frequency domain. The adjusting function had been gained in frequencydomain. Calculate the inverse Fourier Transform of the adjusted function; we get the adjusting function in the spectrum domain. Using the light source and the color filters to test the three methods, we found that they can compensate the attenuation of the optical fiber probe accurately. 二,For the first time the apples and the meat were taken as the object, the chromatic parameters was taken as the basis, the sorts of the apples and the sorts of the meat was classified. Good result has been achieved according the chromatic value through a vast amount of experiments. The four sorts of many samples have been measured with the spectral analysis system. The relative reflective spectrum is defined as a ratio line of the sample's spectrum line versus the spectrum line of the nearby good ones. We discovered that the relative reflective spectrum lines of the fleckless areas are smooth nearby 1. Those of the pushed areas are relatively smooth, but the norms of the lines are more than 1. The pushed areas are smoother than the normal spherical surface that the values of the spectrum lines are higher than those of the nearby areas'. The spectrum lines of the scar areas and the rotten areas are alike. Most of the numbers are under 1.0. The spectrum lines of the scar areas become much smoother at 685nm while those of the rotten areas have a wave crest at 685nm and a trough at 520nm. According to the CIE standard, we calculate the chromatic aberrations of the samples. The chromatic aberration is 32.138, 19.154, 11.564, 4.898 according to the order of the rotten, the scar, the pushed and the fleckless. From the color coordinates, the coordinate x of the rotten and the scar are a lot higher. This is considered as the energy of red light become more. But the change for the pushed is not so manifest. The meat fresh level is divided into the new, the less new and the decay according the national standard. After chromatic calculation of the spectrum data which has been gained through the testing of the standard fresh level meat, the result was obtained as followed: the chromatic aberration(versus the new type) of the new type(versus the new type) is 0-18,the average value is 7.022;Thechromatic aberration of the less new type is 15-29, the average value is 21.661;The chromatic aberration of the decay type is 28-38, the average value is 33.619. It is proved that we can get the meat fresh level coarsely through the chromatic calculation. 三,The paper had put forward a three-stage BP-ANN to setup a relation between the nature of fleck and its reflective spectrum, and a BP neural network was founded to classify the meat fresh levels. The classifying accuracy is above 80%. In the course of apples classifying, a standard three-layer BP-ANN includes an input layer, a hidden layer and an output layer. The number of every layer, the transfer functions of the hidden layer and the output layer and the range of the output are all main parameters for a BP network. Now we take a BP network, which is designed to distinguish the scar from the rotten, for example to tell how these parameters can effect on the veracity of the network (including the precision and extension ability). We choose 16 relative reflective spectrum lines as training samples and another 10 as testing sample set. The relative intension values, at every 5nm from 500nm to 730nm, totally 47 nodes are for the input layer, so the input layer includes 47 nodes. One node for the output layer, its transfer function is linear. Transfer function of the hidden layer is double polarity sigmoid function. After training and changing the parameters, We found that:1. The more nodes in hidden layer, the more accurate the network is. But too many nodes or too many training epochs also debase the accuracy. 2. The smaller the range of output layer is, the lower the differentiate error of both the training samples and the testing samples are, and the lower the resolving power is. When the range is enlarged, the number of training epochs will become larger and the resolving power will be enhanced. Taking the resolving power, the extension power and the constringency rate into account, this paper setup a BP-ANN with 6 nodes on hidden layer and output range [0.1 0.9]. The BP network can identify two models accurately under the samples with 20% noise. It is proved that the network has strong non-linear model-identified power. A three-layer BP neural network was founded to classify the fleckless, the...
|
PDF Full Text Request