JCSE, vol. 20, no. 3, pp.177-188, September, 2026

DOI: http://dx.doi.org/10.5626/JCSE.2026.20.3.177

Real-Time Embedded Image Processing Using Optimized CNN-Transformer Architectures

Jihai Lei
College of New Energy, Longdong University, Qingyang, China

Abstract: In monitoring systems, embedded systems collect real-time data through sensors and cameras for image processing and transmission, but real-time image processing still faces challenges. Therefore, this study designs a real-time semantic segmentation model built on convolutional neural networks (CNNs) and Transformers, and a real-time object detection model on the basis of faster region-based CNN (Faster R-CNN). The results showed that the mean intersection-overunion (mIoU) of the proposed real-time semantic segmentation model was 80.1%, the frame rate was 172.4 FPS, and the image segmentation accuracy was 93.68%. The detection accuracy of the proposed real-time object detection model was 92.17%, and the recall rate was still 90.45%. The average absolute error and F-measure index in the ECSSD dataset were 0.025 and 0.951, respectively. The proposed model converged the fastest, and the error converges to the minimum value when epoch was around 90. The experimental results demonstrated the real-time image processing performance of the research model. The research results contribute to the promotion of computer vision and promote the widespread utilization and development of embedded systems in multiple fields.

Keyword: No keyword

Full Paper:   0 Downloads, 2 View