Machine vision detecting potato mechanical damage based on manifold learning algorithm

Abstract: Buds and uneven surface of potatoes have caused problems to detect the mechanical damage based on machine vision. The lighting conditions and gray value changes of defect region have great impacts on the pixel level feature extraction. While manifold learning methods have been extensively studied in the face recognition, they have not been used for the external quality inspection of agricultural products. The manifold learning method is mainly divided into linear and nonlinear manifold learning algorithms. The nonlinear manifold learning algorithm includes isometric mapping (Isomap), locally linear embedding (LLE), laplacian eigenmaping (LE). The linear algorithm is extension of the nonlinear methods such as principal component analysis (PCA) and multidimensional scaling (MDS). In order to weaken the influence of the buds and uneven surface on potatoes mechanical damage detection, the image was characterized by using low dimensional manifolds. A mechanical damage detection method for potatoes was provided based on manifold learning. In this study, the Saliency and H images were firstly segmented on the potato regional image. The segmentation accuracies of both images are 100%. However, Saliency-H method can the potato's location information of the image by unsupervised pattern was automatically obtained. In addition, Saliency-H method was faster (average elapsed time is 477.7ms) than H method with a high data compression rate. After the potato region images were resampled from 1024×768 to 64×64, the features of potato images were extracted from the resample images by using the three manifold learning methods: principal component analysis (PCA), isometric mapping (Isomap) and locally linear embedding (LLE). Thirdly, the three corresponding SVM classification models were developed based on their features. Finally the parameters of the models were optimized to develop corresponding optimal classification models by using the grid search method (grid search), genetic algorithm (GA) and particle swarm optimization (PSO). The best three classification models were obtained through comparing the recognition results of SVM classification models. Test results showed that the training set recognition rate of PCA-SVM classification model was 100%, the test set recognition rate was 100%. The best parameter optimization method was grid search, the best number of features was 40, the test parameter c was equal to 27.8576 g. The training set recognition rate of Isomap-SVM classification model was 100%, the test set recognition rate was 91.7%, the best parameter optimization method was GA, the best number of features is 4, the test parameter c was equal to 27.8576 g. The training set recognition rate of LLE-SVM classification model was 100%, the test set recognition rate was 91.7%, the best parameter optimization method was PSO, the best number of features is 19, the test parameter c equals 0.1000, g equals 18.8827. These results indicate that potatoes mechanical damage detection is feasible using three manifold learning methods including PCA, Isomap and LLE. PCA-SVM classification model is the best classification model.