Study On Appearance Feature Based Diagnosis For Children With Autism Spectrum Disorder

Autism Spectrum Disorder(ASD)is a neurological development disorder that begins in the early development.Early diagnosis of ASD plays a crucial role in the intervention of ASD,especially for children with ASD.An ASD child needs to be diagnosed by experienced doctors combined with various examinations and assessment,but this method is time-consuming,laborious and susceptible to subjective factors.Therefore,in this paper,we explored two different diagnosis methods for autism based on external characteristics.In the first stage,we proposed an ASD children classification algorithm based on facial expressions around the collected autistic children’s video database(Ext-Dataset,including 136 ASD children and 136 TD children).This method achieves the purpose of ASD diagnosis by fusing features from facial expressions,head pose angle,and head trajectory length.In order to extract facial expression features,we propose a novel facial expression recognition algorithm based on deep learning network.The recognition accuracy of this algorithm on FER2013 and CK+is 68% and 99.2%,respectively.In addition,we apply accumulative histogram to extract the spatial and temporal information of facial expressions,the head pose angle,and the head trajectory length,and send it to LSTM for classification.Finally,we achieved 96.7% accuracy on the Ext-Dataset.In the second stage,we started by supplementing the autistic children database(ACVD)composed of 135 TD kindergarten children on the basis of the database collected in the first stage.First,we extracted the gaze through a gaze estimation network algorithm based on a convolutional neural network,and then the gaze is mapped to the screen and the accumulative histogram is used to further extract features.Finally,by fusing the head pose angle features and sending it into LSTM for classification,the accuracy of 94.8% was finally achieved in the ACVD database.In addition,we use different gaze estimation algorithms or add different noises to analyze the impact on the diagnosis.