Robust stereo disparity estimation plays a critical role in minimally invasive surgery, where dynamic soft tissues, specular reflections, and data scarcity pose major challenges to traditional end-to-end deep learning and deformable model-based methods. In this paper, we propose a novel disparity estimation framework that leverages a pretrained StyleGAN generator to represent the disparity manifold of Minimally Invasive Surgery (MIS) scenes and reformulates the stereo matching task as a latent-space optimization problem. Specifically, given a stereo pair, we search for the optimal latent vector in the intermediate latent space of StyleGAN, such that the photometric reconstruction loss between the stereo images is minimized while regularizing the latent code to remain within the generator’s high-confidence region. Unlike existing encoder-based embedding methods, our approach directly exploits the geometry of the learned latent space and enforces both photometric consistency and manifold prior during inference, without the need for additional training or supervision. Extensive experiments on stereo-endoscopic videos demonstrate that our method achieves high-fidelity and robust disparity estimation across varying lighting, occlusion, and tissue dynamics, outperforming Thin Plate Spline (TPS)-based and linear representation baselines. This work bridges generative modeling and 3D perception by enabling efficient, training-free disparity recovery from pre-trained generative models with reduced inference latency.

Recent advancements in natural language processing have given rise to numerous pre-training language models in question-answering systems. However, with the constant evolution of algorithms, data, and computing power, the increasing size and complexity of these models have led to increased training costs and reduced efficiency. This study aims to minimize the inference time of such models while maintaining computational performance. It also proposes a novel Distillation model for PAL-BERT (DPAL-BERT), specifically, employs knowledge distillation, using the PAL-BERT model as the teacher model to train two student models: DPAL-BERT-Bi and DPAL-BERT-C. This research enhances the dataset through techniques such as masking, replacement, and n-gram sampling to optimize knowledge transfer. The experimental results showed that the distilled models greatly outperform models trained from scratch. In addition, although the distilled models exhibit a slight decrease in performance compared to PAL-BERT, they significantly reduce inference time to just 0.25% of the original. This demonstrates the effectiveness of the proposed approach in balancing model performance and efficiency.

This study addresses the limitations of Transformer models in image feature extraction, particularly their lack of inductive bias for visual structures. Compared to Convolutional Neural Networks (CNNs), the Transformers are more sensitive to different hyperparameters of optimizers, which leads to a lack of stability and slow convergence. To tackle these challenges, we propose the Convolution-based Efficient Transformer Image Feature Extraction Network (CEFormer) as an enhancement of the Transformer architecture. Our model incorporates E-Attention, depthwise separable convolution, and dilated convolution to introduce crucial inductive biases, such as translation invariance, locality, and scale invariance, into the Transformer framework. Additionally, we implement a lightweight convolution module to process the input images, resulting in faster convergence and improved stability. This results in an efficient convolution combined Transformer image feature extraction network. Experimental results on the ImageNet1k Top-1 dataset demonstrate that the proposed network achieves better accuracy while maintaining high computational speed. It achieves up to 85.0% accuracy across various model sizes on image classification, outperforming various baseline models. When integrated into the Mask Region-Convolutional Neural Network (R-CNN) framework as a backbone network, CEFormer outperforms other models and achieves the highest mean Average Precision (mAP) scores. This research presents a significant advancement in Transformer-based image feature extraction, balancing performance and computational efficiency.