Non-contact estimation of Respiratory Rate (RR) has revolutionized the process of establishing the measurement by surpassing some issues related to attaching sensors to a body, e.g. epidermal stripping, skin disruption and pain. In this study, we perform further experiments with image processing-based RR estimation by using various image enhancement algorithms. Specifically, we employ Super Resolution (SR) Deep Learning (DL) network to generate hallucinated thermal image sequences that are then analyzed to extract breathing signals. DL-based SR networks have been proved to increase image quality in terms of Peak Signal-to-Noise ratio. However, it hasn't been evaluated yet whether it leads to better RR estimation accuracy, what we address in this study. Our research confirms that for estimator based on the dominated peak in the frequency spectrum Root Mean Squared Error improves by 0.15bpm for 8-bit and by 0.84bpm for 16-bit data comparing to original sequences if hallucinated frames are used. Mean Absolute Error is reduced by 0.63bpm for average aggregator and by 2.06bpm for skewness. This finding can enable various remote monitoring solutions that may suffer from poorer accuracy due to low spatial resolution of utilized thermal cameras.