Facial Image Inpainting Using Vision-Based Quantized Conditional Generative Adversarial Network (QCGAN) on Edge Device

Image inpainting is considered a task in the field of computer vision aimed at filling in missing or corrupted parts of an image. By using deep learning techniques, they learn a mapping among the distorted image and its corresponding counterpart undistorted image. Applications for conditional image inpainting include enhancing, restoring, and creating new images. The goal of this task is to generate a plausible completion of the missing parts that is consistent with the visual aspect and structure of the surrounding image regions. The conventional method of image inpainting entails using information from the surrounding areas to fill in any damaged or missing portions of images. We proposed a QCGAN by training two neural networks parallelly: the discriminator network, which has a loss of 0.10833, and the generator network, which has a loss of 3.77887. The discriminator network assesses the generated pictures to determine if they are real or fraudulent, and the generator network attempts to generate a complete image of an incomplete image. Through this process, the generator network learns to generate realistic and accurate images that can rectify or restore incomplete or impaired portions of an image. Experiments using a customized dataset demonstrate the stability and effectiveness of our suggested model, as well as its higher naturalness, semantic similarity, and validity in terms of visual experience. The model underwent post-dynamic range quantization, and its size was reduced by 74.8% to enable deployment on edge devices. We achieved the PSNR value of 30.76 and the SSIM of 0.86.