The use of Artificial Intelligence (AI) in Internet of Things (IoT) ecosystem has been growing exponentially, enabling various Computer Vision (CV) applications. These applications must handle large image data demanding reliable communication systems that retain image quality for downstream Deep Learning (DL) tasks. Existing communication systems, such as Orthogonal Frequency Division Multiplexing (OFDM), promise improvements in data rate, spectral efficiency, and mitigation of multipath fading; however, these systems often distort the received images due to complex channel environments and impairments from various physical layer (PHY) blocks. Source Coding is one such PHY block, which aims for compression savings at the expense of image quality. Therefore, in this study, we evaluate the performance of a DL model for downstream image recognition tasks, where images are transmitted over communication systems utilizing various source coding schemes over complex channels. Experimental analysis shows that Variable-Length Coding (VLC) retains superior image quality, which results in over 95% DL model accuracy throughout the experiment.