| Popis: |
Abstract Autism spectrum disorder (ASD) is a complex developmental issue that affects the behavior and communication abilities of children. It is extremely needed to perceive it at an early age. The research article focuses on attentiveness by considering eye positioning as a key feature and its implementation is completed in two phases. In the first phase, various transfer learning algorithms are implemented and evaluated to predict ASD traits on available open-source image datasets Kaggle and Zenodo. To reinforce the result, fivefold cross-validation is used on the dataset. Progressive pre-trained algorithms named VGG 16, VGG 19, InceptionV3, ResNet152V2, DenseNet201, ConNextBase, EfficientNetB1, NasNetMobile, and InceptionResNEtV2 implemented to establish the correctness of the result. The result is being compiled and analyzed that ConvNextBase model has the best diagnosing ability on both datasets. This model achieved a prediction accuracy of 80.4% on Kaggle with a batch size of 16, a learning rate of 0.00002, 10 epochs and 6 units, and a prediction accuracy of 80.71% on the Zenodo dataset with a batch size of 4, a learning rate of 0.00002, 10 epochs and 4 units. The accuracy of the model ConvNextBase is found challenging in nature as compared to an existing model. Attentiveness is a parameter that will accurately diagnose the visual behavior of the participant which helps in the automatic prediction of autistic traits. In the second phase of the proposed model, attentiveness is engrossed in identifying autistic traits. The model uses a dlib library that uses HOG and Linear SVM-based face detectors to identify a particular facial parameter called EAR and it is used to measure participants' attentiveness based on the eye gaze analysis. If the EAR value is less than 0.20 for more than 100 consecutive frames, the model concludes the participant is un-attentive. The model generated a special graph for a time period by continuously plotting the value of EAR based on the attention level. The average EAR value will depict the attentiveness of the participant. |
Full text from SpringerLink