Performance Evaluation of Artificial Neural Network Methods Based on Block Machine Learning Classification

Traditional (pixel-by-pixel) classification techniques are time-consuming, whereas semantic segmentation in machine learning requires assigning class labels to each pixel in an image. This study proposes a block-by-block (5x5 chunks) segmentation method for semantic segmentation, which involves image dissection, feature extraction, and model training based on specific color and textural properties. Thirty cat photos from the Oxford-IIIT Pet dataset were used for evaluation. Five different Artificial Neural Network (ANN) models, including LM, BGFGS, RP, SCG, and GDX, were trained and assessed for both pixel-based and block-based methods. The accuracy of the block-based classification ranges from 82.94% to 85.83%, surpassing the pixel-based approach, which ranges from 70.82% to 76.47%. The processing time for the models also improved with the block-based method. For the pixel-based approach, RP model takes the longest processing time i.e., 242.39 seconds, while GDX model takes the shortest processing time i.e., 49.89 seconds. For the block-based approach, LM model takes the longest processing time i.e., 13.86 seconds, while GDX still has the shortest processing time i.e., 5.98 seconds. Therefore, block-based methods can be seen as more efficient and accurate for classification models. The LM model achieved the highest accuracy on test images, ranging from 94.72% to 89.81%, while the GDX model had the lowest accuracy, ranging from 92.96% to 81.15%. The remaining models, RP, SCG, and BFG, have intermediate levels of accuracy.