Wild horseshoe crab image denoising based on CNN-transformer architecture.
Journal:
Scientific reports
PMID:
40185943
Abstract
The natural habitats of wild horseshoe crabs (such as beaches, shallow water areas, and intertidal sediments) are complex, posing challenges for image capture, which is often affected by real noise factors. Deep learning models are widely used in image denoising techniques, including methods based on CNNs and Vision Transformers (ViT), each with its own advantages and disadvantages. In this paper, we construct a dataset of wild horseshoe crab images and propose a CNN-Transformer hybrid model that combines multi-head transposed attention mechanisms, gating mechanisms, and depth-wise separable convolution to reconstruct key features of wild horseshoe crab images. The model uses a linear complexity multi-head transposed attention mechanism applied to the channel dimension and combines it with gating mechanisms and depth-wise convolutions to reconstruct contextual features, fully leveraging global contextual relationships across feature dimensions to optimize denoising quality. Extensive experimental results show that the model can accurately restore key features of wild horseshoe crab images, which is of great significance for their tracking and localization.