Image denoising is a fundamental task in computer vision, serving as a preprocessing step to enhance image quality and improve the performance of subsequent recognition systems. Its importance is particularly evident in Internet of Things (IoT) applications, where sensors frequently capture noisy data. While autoencoder-based denoising models have achieved strong performance, their evaluation on resource-constrained embedded platforms with real on-device measurements remains largely unexplored. This study addresses this gap by presenting one of the first feasibility investigations of deploying an image-denoising autoencoder on a microcontroller unit (MCU). We introduce TinyCDAE, a deployment-oriented lightweight convolutional denoising autoencoder designed for resource-constrained microcontrollers, and demonstrate its deployment on ESP32. The model is first evaluated on a PC environment and compared with a baseline fully connected denoising autoencoder (DAE). The results show that TinyCDAE, with only 373 parameters compared to 52,064 for the baseline DAE, significantly outperforms the baseline in terms of PSNR and SSIM across multiple noise levels. TinyCDAE is then deployed on an ESP32 microcontroller, achieving real-time denoising with an average inference latency of 133.57 ms per image. Beyond demonstrating technical feasibility, our on-device measurements on ESP32 confirm that a compact denoising autoencoder can operate effectively on resource-constrained IoT devices.