An Emotion being a complex psychological state that involves both experience and action becomes a challenge to be recognized accurately by programmable codes. This paper demonstrates the method of identifying each out of seven basic emotional states (happiness, surprise, fear, anger, fear, disgust, sadness and neutral) and characterizing them as either positive or negative (valence) from images in a given dataset. This has been achieved to a higher accuracy by a Convolutional Neural Network designed with Double Average filters and the SELU (Scaled Exponential Linear Unit) activation units. The images from the FER 2013 dataset is processed (converted to gray scale and the dimensions set to 48x48) and given as input to the CNN. The Double Average Filters remove the noises much more efficiently than Average Filters, since the process is repeated to give even lesser intensity variations between the pixels. The SELU activation used in the CNN gives an internal normalization on the filtered images, which results in a much better identifying of emotions than with other activation unit. The SELU in recent times, as mentioned by other researchers too, is a promising part of any networks that can be used in Machine Learning. The proposed novel CNN model has a training accuracy of more than 96.53%.