Automatic recognition of complete palynomorphs in digital images

Images of dispersed kerogen preparation are analysed in order to detect palynomorphs of elliptical/spherical shape. This process consists of three automatic stages. Firstly, the background of the image is segmented from the foreground. Secondly the foreground particles are segmented into individual regions. Finally a trained classifier is used to label a region as either containing a palynomorph or containing other material. Ten classifiers were trained and then tested using a ten times tenfold cross-validation. Typically the number of regions in the image containing other material exceeds by far the number of regions with palynomorphs. Hence the problem of imbalanced classes was addressed. Training data was sampled ten different times maintaining a balanced class distribution. Thus the accuracy for each classifier was assessed on 1,000 testing sets. The logistic classifier was chosen and a certainty threshold was selected by ROC curve analysis. The final automatic recognition has accuracy of 88%, sensitivity of 87% and specificity of 88%.