Research On The Non-destructive Testing Of Pork Freshness Based On Multi-information Fusion

Chinese people consumed the most of pork in the world, and China is one of the most important pork export country. Freshness is one of the most important indexes to evaluate pork quality.The spoilage pork would lead to food poisoning. It is important to enhance hygienic examination to ensure the safety of pork products. Therefore, built a fast and accurate method to evaluat pork freshness has a scientific and practical value.Porky loin and porky hind leg were selected as test materials.Computer vision technology and electronic nose technology used to got pork image information and odor information. In order to improve the recognition rate of pork freshness, multi- information fusion was adopted to fuse pork image feature information with gas feature information and a non-destructive testing model of pork freshness was established. The main conclusion as follows:1. Establish pork image acquisition system by computer vision technology. Image sensor and image acquisition card were selected in accordance with the need of pork image acquisition. An image acquisition device was established.2. Optimize the image noise elimination method and selected pork image feature information. Establish a pork freshness model based on computer vision technology.Compared the effect of median filter and mean filter method on pork image noise elimination, took median filter as the best method for pork image noise elimination. adopt RGB,L * a * b *,HIS color model parameters and pork samples storage time to establish fitting model, according to the coefficient value R2 of the fitting model ,selected L *,a *,b * value as the best pork image characteristics .The TVB-N value of the training set and validation set samples were measured via Semi-micro Kjeldahl method. All the samples were coded into three groups based on the TVB-N value. The pork freshness detection model was established based on pork color characteristic parameters. Training set which contain 394 pork samples was adopted to train the model. Freshness code and image characteristic parameters were adopted as network input. The validation set which Contain 96 pork samples were adopted to verify the model. The result indicated that: this model can identify loin samples from the hind leg samples, the classification accuracy rate is 100%, and however the correct classification rate of pork freshness was 76%. It was found that the model based on pork color characteristic parameters are effective to distinguish different position of pork while the capability of pork freshness prediction is poor.3. Pork odor acquisition system design including tow parts: The hardware design and software design. Select TGS822, TGS832, TGS825 and TGS826 to build up gas sensor array for acquisition of pork odor. Design odor acquisition software based on LabVIEW.4. The test parameters of Electronic nose system: head space of time, the sample weight, and sampling time were optimized. Took BP neural network and adopted odor characteristic parameters to build a pork freshness identification model.Adopted PCA and curve fitting as the analysis method, the optimal measurement condition for electronic nose are as follows: headspace for 3min, 60g for sample weight, 90s for sampling time. Subtract gas sensor voltage steady-state value from initial value, and the difference value was adopted as odor characteristic parameters.Code the pork sample freshness according to their TVB-N value. Training set which contain 394 pork samples was adopted to train the model. Freshness code and odor characteristic parameters were adopted as network input. The validation set which Contain 96 pork samples were adopted to verify the model. The test accuracy of the model is 86.5%.It was found that the model based on pork odor characteristic parameters are effective to distinguish pork freshness, but it couldn't distinguish hind leg samples from loin samples.5. Took multi-information fusion technology to fuse the odor information and image information. A pork freshness identify model was established based on multi-information fusion technology. Compared pixel-level,feature-level and decision-level fusion mode, the result indicated that feature fusion mode was the most suitable mode for fusion pork image information and odor information. Code the hind leg samples and loin samples freshness according to their TVB-N value.Took BP neural network,RBF neural network and Least Squares Support Vector Machine to built a pork freshness identify model. Training set which Contain 394 pork samples were adopted to train the BP neural network. Samples freshness code> odor characteristic parameters and image characteristic parameters was adopted as network input. The validation set which contain 96 pork samples was adopted to verify the model. The result indicated that: all the multi-information fusion model could distinguish hind leg samples from loin samples in the validation set .The classification accuracy rate is 100%, and however the correct identification rate of freshness to the BP neural network,RBF neural network and Least Squares Support Vector Machine model were 91.67%,94.79%and89.6% respectively.The results demonstrated that compared with the model based only on image sensor characteristic parameters or gas sensor characteristic parameters, the multi-sensor fusion model had more accuracy and applicability. The Correct identification rate of freshness result indicated that: RBF neural network is the best method to build multi-information fusion pork freshness identify model.