<p>The recent advancement in the AI technology serves the needy people to uplift their survival during their lifetime. It not only helps them to cope up with the situation but also assist the care takers who supports them. The non-verbal expression especially facial expression plays a major role in social interaction to communicate effectively. The facial expression given by the common people can easily be identified. However, it is quite difficult for the people with Autism Spectrum Disorder (ASD). Thus, an attempt is made in this paper to recognize the facial expression of children with ASD through deep learning methodologies. A novel Inter-Channel Attention Network (ICAN) is proposed in this paper with three well-defined phases viz., separable convolution, inter-channel attention and squeeze excitation. Since the horizontal and vertical orientation of the facial region contributes for emotion identification, two separable convolutions are employed in the first phase. The outcomes of these separable convolutions are processed independently with subsequent operations in the next two phases and concatenated in the recognition phase. In the second phase, the significance of the feature maps is thoroughly analyzed in the Inter-Channel Attention Module (ICAM) and separated into two sublists viz., most dominant channels (MDC) and least dominant channels (LDC). The feature maps of these channels are enhanced independently and merged together. Then, this reduced collection of feature maps is applied with squeeze excitation in the last phase which results in enriched collection of feature maps. Finally, they are fed into the recognition phase. Through rigorous experimentation and validation across diverse datasets, the proposed model demonstrates promising results by training the model with regular dataset and tested with ASD dataset. It achieves the accuracy of 89.45%, 85.68%, and 74.87% respectively for the experiment setup viz., RAF DB-ASD, FER 2013-ASD and FED RO-ASD.</p>