Hardware Implementation of a Deep Learning-based Model for Image Quality Assessment

Abstract

Image quality assessment is very important for accurate analysis and better interpretation. In reality, environmental effects and device limitations may degrade image quality. Recently, many image quality assessment algorithms have been proposed. However, these algorithms require high computation overhead, making them unsuitable for mobile devices, such as smartphones and smart cameras. This paper presents a hardware implementation of an image quality assessment algorithm based on a Lightweight Convolutional Neural Network (LCNN) model. Many advances have been made in the construction of high-accuracy LCNN models. The current study used EfficientNet V2. The model achieved state-of-the-art image classification performance on many famous benchmark datasets while having a smaller size than other models with the same performance. The model was utilized to learn human visual behavior through understanding dataset information without prior knowledge of target visual behavior. The proposed model was implemented employing a Field Programmable Gate Array (FPGA) for possible integration into mobile devices. The Xilinx ZCU 102 board was implemented to evaluate the proposed model. The results confirmed the latter’s efficiency in image quality assessment compared to existing models.