Adaptive Multi-Scale Image Stitching Using an Attention-Enhanced BiFPN With Contrast-Aware Optimization

Abstract

To address the challenges of low stitching accuracy and limited robustness in complex scenes, this study proposed an image stitching model based on an improved Bi-directional Feature Pyramid Network (BiFPN). The model enhances performance through three key optimizations. First, an adaptive weighting mechanism dynamically balances the global and local contributions of multi-scale features. Second, a Squeeze-and-Excitation (SE) attention mechanism strengthens feature extraction in critical stitching regions such as edges and textures. Third, a global contrast enhancement module mitigates illumination variation effects on feature matching through multi-scale histogram equalization and adaptive calibration. Experiments were conducted on two benchmark datasets: Microsoft Common Objects in Context (MS COCO) and the Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI). From MS COCO, 1,500 image pairs were selected (500 with illumination variations and 500 with scale variations). From KITTI, 1,500 image pairs were selected (800 static scenes and 700 dynamic targets). Each dataset was split into training and validation sets with an 8:2 ratio. Training used a batch size of 16, 50 epochs, and an initial learning rate of 0.001 with a 50% decay every 10 epochs. Comparative methods included traditional algorithms such as Oriented FAST and Rotated BRIEF (ORB) and Scale-Invariant Feature Transform (SIFT), as well as deep learning approaches including Vision Transformer-Large/16 (ViT-L/16) and the Stitch Generative Adversarial Network. The proposed model outperformed all baselines in complex scenarios. On the MS COCO dataset with illumination variations, the mean squared error (MSE) reached 1.12×10⁻²—69.09% lower than ORB and 39.46% lower than ViT-L/16. The peak signal-to-noise ratio (PSNR) increased to 34.89 dB, improving by 5.11 dB over SIFT and 2.75 dB over other models. The structural similarity index (SSIM) reached 0.946, exceeding competing methods by 7.26%. On the KITTI dataset with dynamic targets, the feature matching accuracy reached 92.3%, a 17.95% improvement over SIFT, while the stitching time decreased to 1.78 s, 30.47% faster than other models. The model maintained high robustness under parallax and motion blur conditions, providing precise and efficient image stitching for vision-based control and automation tasks such as robotic navigation and industrial monitoring.