A Novel Approach With Enhanced Accuracy and Efficiency for Image Segmentation

• Published in: IEEE access Vol. 13; pp. 63587 - 63599
• Main Author: Liang, Jinliang
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Adaptation models; Attention mechanisms; Biomedical imaging; Computational efficiency; Computational modeling; Computer architecture; Computer vision; Deep learning; Efficiency; Entropy; Image analysis; Image segmentation; Machine learning; Medical imaging; Pixels; Satellite imagery; Structural integrity; Training; U-Net
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2025.3552588

Image segmentation is a fundamental and critical task in computer vision, with applications across various domains, such as medical imaging, autonomous driving, and satellite image analysis. However, existing methods often face challenges in achieving both high accuracy and computational efficiency simultaneously. In this paper, we propose a novel approach to image segmentation that addresses these challenges. Our method is based on the sophisticated U-Net deep learning architecture which integrates multiple advanced techniques. We introduce residual connections to mitigate the vanishing gradient problem, thereby enhancing the stability and convergence speed of the training process. Dilated convolutions are incorporated to expand the receptive field of the network without significantly increasing the number of parameters, enabling the model to capture more contextual information. Additionally, attention mechanisms are integrated to allow the model to focus its computational resources on the most relevant and salient regions of the image, resulting in more precise segmentation. A unique multi-scale feature fusion strategy is employed to combine features from different levels of the network, enriching the feature representation and improving the model's ability to handle objects of various sizes. Furthermore, we design a novel loss function that combines traditional cross-entropy loss with a structural similarity index measure (SSIM) loss. The cross-entropy loss ensures accurate pixel classification, while the SSIM loss preserves the structural integrity and consistency of the segmented images. Extensive experiments on several benchmark datasets, including Pascal VOC 2012, Cityscapes, and COCO-Stuff, demonstrate that our method outperforms numerous state-of-the-art approaches in terms of both segmentation accuracy and computational efficiency. The results show that our method achieves higher mean intersection over union (mIoU), better pixel accuracy, and improved boundary F-score compared to competing methods. Additionally, our approach exhibits greater robustness in handling noisy and occluded images, further validating its practical applicability in real-world scenarios. Overall, our novel approach represents a significant advancement in image segmentation and holds great potential for a variety of practical applications.