Research on an Improved Unsupervised Algorithm for Wood Image Stitching

This study proposes a strategy based on pre-training and fine-tuning aimed at enhancing model performance and training speed. Initially, the model is pre-trained on natural images to learn a universal feature representation. Subsequently, it undergoes fine-tuning on a wood dataset to acquire specific characteristics of wood images. The network reconstruction phase is designed with dual branches: a low-resolution branch and a high-resolution branch. The low-resolution branch integrates a lightweight channel attention mechanism, autonomously learning the significance of each channel. The high-resolution branch leverages an enhanced self-attention mechanism to model the remote dependencies present in the input data. The reconstruction network incorporates multiple residual structures, effectively overcoming gradient vanishing and explosion challenges. Experimental results indicate that the proposed approach achieves a harmonious balance between computational efficiency and output quality, proving effective in the stitching of wood images.