Development of PCAHOG feature extraction with ROTN-ELMSOM classifier for Aspergillus species identification

Aspergillus is one of the most ubiquitous of the airborne saprophytic fungi, capable of thriving in various climatic conditions and causing multiple types of illness. These fungi can be both beneficial and harmful to humans and animals. Direct microscopic is often used by microscopist as an alternative for identifying any specimens suspected of fungal infection. However, confirmation towards identification is often necessary due to the complex and dissimilar structure of Aspergillus in each cycle. In addition, misidentification can occur, especially among species with similar features. To prevent misidentification and the need of accurate identification, a computer-based Aspergillus species identification approach is proposed which encompassed recognition and identification stages. The recognition process the crucial stage to detect the presence of fungi therefore, an active region-based segmentation method is proposed. This method is not solely depending on the gradients or sharp edges of the object and implementing level set function for curve evolution which effectively reduced computational cost. Two different methods were tested and compared, aiming to observe the approaches’ ability to segment different 240 of Aspergillus images representing four species. Experiments conducted have been compared with the baseline technique and the proposed method is outperformed in terms of accuracy and specificity with an average of 90% and PSNR value of greater than 40dB. Meanwhile the active contour (Snake) was slightly underperformed but well performed in terms of sensitivity with values exceeding 80% for all the species. Afterwards, moving on the classification stage, principal component analysis and histogram of gradient oriented (PCAHOG) features extraction methods are employed in conjunction with semi-supervised robust optimized threshold neural network extreme learning machine and self-organizing map model (ROTN-ELMSOM) classifiers. PCAHOG captured pixels gradient values to reduce the dimension of the database while retaining the valuable data points. The resulting compact data representation is subsequently utilized in ROTN-ELMSOM classification and capitalized on the advantages of both unsupervised and supervised classifier along with an adjustment in the hidden layer. The SOM approach facilitated learning the underlying structure thus, enabling ELM approach to execute the classification process. However, to address the overfitting due to limited data availability, a threshold parameter was introduced in the hidden layer. This addition enhanced the classification performance by minimizing the inclusion of redundant neurons. Furthermore, a few other experiments are conducted to determine optimum values for parameters such as number of epochs, neighbourhood size, number of neurons, training ratio and weight rule respectively for ROTN-ELMSOM classification process. Compared to original ELMSOM, the acquired accuracy, sensitivity and specificity that are greater than 96% substantiated the effectiveness of the proposed PCAHOG features extraction and ROTN-ELMSOM classification methods.