Recognizing Daily Activities of Children with Autism Spectrum Disorder Using Convolutional Neural Network Based on Image Enhancement

Independence for individuals with disabilities, Children with Autism Spectrum Disorder (ASD), need skills to perform daily activities. This study focuses on recognizing the daily activities of children with ASD using a Convolutional Neural Network (CNN) based on augmented images. The CNN architectures employed are Visual Geometry Group 19 (VGG19) and MobileNetV2, while image improvement techniques include Histogram Equalization, Contrast Stretching, and Contrast Limited Adaptive Histogram Equalization (CLAHE). The data consists of eating (606 videos) and drinking (477 videos) activities recorded by therapists or parents. CLAHE proved the most effective, achieving an SSIM of 0.998 and a PSNR of 38.466 for the eating activities, an SSIM of 0.998, and a PSNR of 38.296 for the drinking activities. Experimental results using CLAHE and VGG19 showed a recognition model accuracy of 85%, while VGG19 without image enhancement achieved an accuracy of 83%. CNN with image enhancement achieves slightly better accuracy, though the difference is insignificant.