The Research Of Meat Spectrum Recognition Based On Neural Network And Wavelet

It is a necessary important link to automatically identify the merchandise in the modern logistics. To some special merchandise like kinds of meats, it is difficult to identify species only by naked eyes for their almost similar colors. Besides the religion (the Islamite doesn't eat pork), the healthy concern and the consideration of economic benefits, the research of species recognition becomes more and more valuable.The paper studies the way of species recognition (beef, pork and chicken). Measure their reflective spectra, calculate their chromatic coordinates and tristimulus values, introduce the wavelet neural network to set up a model between the reflective spectra and the characteristics of the meat.A spectrum analysis system is designed to measure the spectrum of reflection of meat surface. The 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).An optic fiber 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. It is so small and very fit for measure on micro areas. A hologram grating controlled by a stepping motor, works as light-splitting system. The showing error is smaller than 0.5nm which proves that the system answers for the CIE's demands. Photoelectric multiplier tube, the photo electronic transferring system, is capability of measuring weak signals. The data-processing software is programmed with Matlab languages. Its functions are normalizing data, calculating chromate, developing surface, building ANN (Artificial Nerve Networks). All of these make up of a function all-ready and structure fast spectrum analysis system.This spectrum analysis system is used by a variety of three meat for the measurement. Relative reflectance spectroscopy is the ratio of the reflectance spectra for the meat and their reflectance spectra for the whiteboard. There was a reception to the many peak spectral signal or burr-like protuberances which can be seen from the lines map. Namely, random noise is caused by the existence of multiple sources of noise. The curve is relative gather in the 665-780nm wavelength interval. Overall, the curve shape is closer between the beef and pork. Relative to the pork and beef, the curve of chicken samples is more intensive and the trend is also significant differences. In addition to the rapid decline of the singular point near 550 nm, the other band appears more smooth upward trend. It appears that it is difficult to identify the three meat from the observation of the original image alone.For this reason, the paper presents a recognition measure based on a wavelet theory and neural network. It combines the partial nature of the wavelet and the automatic recognition of BP artificial neural networks, which has a strong approximation capability. In the classification and recognition of spectral signal, the wavelet transform is used to extract the useful signal as a signal smoothing processing method, then put the extracted information into neural network classification.The main study of the paper include two parts of the research and the realization of wavelet algorithm and the feed-back propagation neural network (i.e. BP net) algorithm.The research of the algorithm includes the following. (1) Signal's WT. A mother wavelet is selected and decomposing level N is decided, then the signal is decomposed by the mother wavelet to N level. (2) The threshold processing of high frequency coefficients. Different thresholds are selected for the high frequency coefficients from 1 to N level. (3) Signal's restructure. The signal is restructed by the low frequency coefficients on N level and high frequency coefficients after processing from 1 to N level.Attention should be paid to the comparability between signal and mother wavelet as well as the size of operation quantity when choosing mother wavelet. The decomposing level is affected by two factors, one is the wavelet type, the other is data length. If the decomposing level is too low, the program runs fast, the time taken is short, but relatively more noise will remain leading to low SNR improve and unsatisfactory filtering result. If the decomposing level is too high, some useful signals and the local characteristics of them will be filtered although most noise is filtered, and the SNR improve will decrease. After a lot of simulation, we have used the actual type of four storeys of db5 wavelet decomposition algorithm.BP neural network usually is the multi-layer neural network based on the back-propagation algorithm (BP net). The inherent characteristics or essential features of samples are more accurate grasped by training to a large number of samples used artificial neural network. Although, this feature won't shown by the formula or charts which we are familiar with. However, these characteristics may be hidden in standard weight and the effective of nonlinear classification is used by the weight standards. The build of BP artificial neural network need to consider the following parameters : input data and input nodes, the number of hidden layer and the number of hidden layer neurons, output range of output neuron and the corresponding levels of transfer function and neural learning. In addition, it need to consider the accuracy and generalization of the network.Through a series of analysis and experiments, we eventually determined to use the 4-BP neural network, 24 for the input layer neurons, 17 for the first hidden layer neurons, the transfer function for the S-tangent function, 4 for the second hidden layer neurons, S-type function for its transfer function, 1 for the output layer neurons, linear transfer function for its transfer function, iterative cycle training for its network training. The detection thresholds are 3.0 and 5.0.The beef is the output of less than 3.0 When the chicken is the out put of larger than 5.0, the pork is the output of between 3.0 and 5.0. After 103 cycles training, we achieved the pre-precision network.As mentioned above, the paper is used the collection and analysis systems of the fiber probe and based on the study of wavelet analysis and neural networks, we proposed a pretreatment neural network recognition algorithm based on wavelet. The simulation results show that this method can greatly improve the recognition of the validity and accuracy of meat and achieve a better discernment.