Performance Analysis of H2O and H2O with HCl Material Image Classification Using Inception V3, VGG19, DenseNet201, and Otsu Segmentation

Challenges in classifying signals with fluctuations remain a focus in the field of image and signal processing. Deep learning technology, especially CNN (Convolutional Neural Network), has proven effective for complex visual classification; however, its performance can still be improved, particularly for signal nonlinearity distributions that are not evenly distributed. This study develops a system for classifying signals that exhibit high fluctuations using a merged Otsu segmentation and deep learning ensemble approach with InceptionV3, VGG19, and DenseNet201 models. The methodology employed is a quantitative study based on a deep learning ensemble. H?O and H?O with HCL signal datasets were processed using Otsu segmentation and then extracted using three CNN architectures, which were then combined with the methods of soft voting and stacking. Evaluation is conducted through the analysis of accuracy, precision, recall, loss, and a confusion matrix. DenseNet201 records the highest accuracy of 95%, precision of 0.90, recall of 0.86, and f1-score of 0.95. InceptionV3 achieves equivalent accuracy (95%) but with a recall of 0.83. VGG19 noted an accuracy of 91%, a precision of 0.82, and a recall of 0.78. The ensemble results show improvement in stability classification, especially in class H?O segmentation. However, the classification class HCL segmentation still shows more mistakes. The integration of Otsu segmentation and deep learning ensemble models has been proven effective in increasing the accuracy of classifying signal fluctuations. Segmentation helps highlight the importance of spatial features, while ensemble enhances model generalization. Research furthermore recommended exploring method segmentation and adaptive data augmentation to handle more complex and unbalanced distributions.