Abstract:
Autism Spectrum Disorder is essentially a condition that disrupts a child’s development, affecting communication skills,
understanding, learning, and interaction difficulties. This study aims to combine LBP (Local Binary Patterns) and HOG (Histogram of
Oriented Gradients) as feature extraction methods in a CNN classifier to improve its accuracy for Autism detection based on digital
facial image. The process followed in this research involves sequentially applying LBP, preprocessing, applying HOG, hyperparameter
tuning, training, testing, and evaluation. The results show that combining LBP and HOG feature extraction methods yields an accuracy
of 80%, which is better compared to using only one of the feature extraction methods: LBP (78%) or HOG (75%), and also better than
not using feature extraction at all, which results in an accuracy of 76%. The combination of feature extraction methods, as done in other
research, is less appropriate in this case, as it only achieved a maximum accuracy of 66%. Therefore, feature extraction combinations
must be selected carefully to avoid making it difficult for the CNN to recognize patterns in facial images. This study concludes that
applying LBP as preprocessing and HOG as feature fusion on a digital image dataset can help improve CNN accuracy for early autism
detection.
