The processing of face information relies on the quality of data resources and therefore the dataset is crucial for image processing. While considering behavioural investigations of facial expression recognition (FER) in autism spectrum disorders (ASD) provide conflicting findings due to its difficulty in processing real-world application domains. Also, the significant intra-class heterogeneity makes the FER a challenging task. Hence, this research introduces a novel facial expression database of ASD children. The database is gathered from ASD children aged 6 to 14 with various levels of disease severity, wherein 45% of the children have severe ASD, 35% have moderate, and 20% have low severity of ASD. A total of 81 Male and 32 female children participated in taking the images of facial expressions. Four expressions are considered when generating the dataset: Happy, Sorrow, Neutral, Angry or Disgusted. The generated dataset is validated by analyzing the facial expression recognition using the deep convolutional neural network (DeepCNN). The accuracy accomplished by the DeepCNN using the proposed ASD facial expression recognition database is 96.14%.