IMPROVING PERFORMANCE OF NEURAL NETWORK CLASSIFERS BY INPUT DATA PRETREATMENT WITH PRINCIPAL COMPONENT ANALYSIS

Authors: Chen, Yud; Nguyen, Minh; Park, Bosoon

Submitted to: Journal of Food Process Engineering
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 3/13/1998
Publication Date: N/A
Citation: N/A

Interpretive Summary: When designing an inspection or grading system to be used on-line at processing plants, one of the major concerns is to design a system with a classifier that is fast, robust, and easily trained. Previously, we reported that with a visible/near-infrared spectrophotometer system, wholesome and unwholesome poultry carcasses could be classified at the current highest processing speed of 90 birds/min with high accuracies. In this paper, we further showed that using principal component analysis (PCA) technique, more accurate and faster processing neural network classifiers could be designed. This paper showed that when sensing moving poultry carcasses in a dark environment, the optimal classifiers for poultry carcasses moving at a speed of 60 or 90 birds/min achieved 100% classification accuracies with PCA pretreatment. When sensing in room light, the best model with PCA pretreatment had either slightly better or comparable accuracies when compared to the classifiers without pretreatment. Furthermore, with PCA pretreatment, the neural network classifiers processed spectral information faster, required fewer training samples, and reduced training time. The results of this study are very valuable to the researchers who are developing and designing real-time on- line machine vision systems for grading or inspecting food and agricultural products. These results are also very useful to FSIS scientists and administrators who are interested in implementing a machine vision system to inspect poultry carcasses at a shackle speed beyond the current highest speed of 90 birds/min.

Technical Abstract: This paper reports the results of applying principal component analyses (PCA) of spectral reflectance data to reduce numbers of input nodes to neural networks for classification of wholesome and unwholesome poultry carcasses. The results showed that, using principal components, the number of input nodes for a backpropagation neural network could be reduced without compromising its accuracy. Models with PCA pretreatment of input data performed better than those without pretreatment. When sensing moving poultry carcasses in a dark environment with a visible/near-infrared spectrophotometer, the neural network classification models with PCA pretreatment achieved 100% accuracies for training, validating, and testing. For carcasses moving at 60 birds/min, 50 factors were required for perfect classification, while for 90 birds/min 30 factors were required. When sensing in room light, the best model was generated with 30 factors for a shackle speed of 60 birds/min, with a test set accuracy of 95.6%. For 90 birds/min, the best model with a test set accuracy of 96.8% was obtained when 15 factors were used. This study showed that PCA was a powerful technique in extracting important features of input data. It reduced the number of input nodes to the neural network classifiers and, in most cases, improved the model's classification accuracy. It also required fewer training samples and reduced training time.