Design and Analysis of Product Visual Communication Systems Based on Big Data Computing Frameworks

Since deep convolutional neural networks (DCNNs) require a large number of parameters when processing high-resolution images, computing resource requirements increase significantly, resulting in long training time and low efficiency. In addition, the generative adversarial network (GAN) often generates blurred and distorted images due to gradient vanishing and mode collapse. To address these issues, this paper combines DCNN and GAN to design and analyze product visual communication in the Hadoop ecosystem, aiming to solve the problems of high model complexity and difficult cross-domain integration in product visualization. The study used model pruning, quantization technology and knowledge distillation to remove redundant network connections and unimportant neurons, convert floating-point operations into low-precision integer operations, reduce model parameters, and improve training efficiency. StyleGAN was used to improve GAN, spectrally normalize the discriminator network in GAN, and introduce a multi-scale loss function to avoid image blur and distortion. Model training is performed in the Hadoop ecosystem. This paper uses the deep learning framework integrated with Hadoop, combined with Spark memory computing, adopts progressive training, applies a multi-stage optimization strategy, introduces adversarial training, and optimizes the model based on user feedback. Experiments show that the model has a PSNR of 29.9 dB, an SSIM of 0.86, an MAE of 7.6, a GAN stability of 7.4, a bending error of 0.4, an FID of 28.1 and an execution efficiency of 0.87. It is superior to other common product visual communication design (VCD) models, verifying the superiority of the product visual communication model based on DCNN and GAN in the Hadoop ecosystem.