Augmented reality requires integration of real-world objects, gestures and actions with the virtual objects. Effective and realistic integration requires solving the complex task of recognition, classification, tracking of objects, gestures and actions, where gesture recognition and action mapping is an active problem in the field of augmented reality, seeking attention for optimized latency, power and throughput. This paper introduces the technique of frame processing with active tile identification to optimize the latency of Convolution neural network in the light of action mapping in augmented reality. The effectiveness of the technique being introduced is evaluated by applying it to the Bharatanatyam Mudra classification and measuring the obtained latency, power and throughput and comparing the obtained results with that of the traditional Convolution neural network. The comparison shows the technique to be effective in terms of the latency, with the best effectiveness factor of 2.30 and least being 1.25.