SIAVC: Semi-Supervised Framework for Industrial Accident Video Classification
Zuoyong Li, Qinghua Lin, Haoyi Fan, Tiesong Zhao, David Zhang

Project

• We propose a Video Cross-set Augmentation Module (VCAM) by incorporating high-confidence unlabeled samples into the augmentation queue and generating various pseudo-label samples through interpolation, to mitigate the sampling experience mismatch while expanding the training data.
• We propose a Super Augmentation Block (SAB) that adds random mask and Gaussian noise to frames based on historical losses to re-augment high-confidence samples. SAB allows for the re-utilization of these strongly augmented samples for better consistency regularization..
• We propose a multi-class accident video dataset called Express Center Accidents 9 (ECA9). ECA9 comprises nine typical accidents in hub-level express processing centers, and provides video-level labels and frame-level anomaly labels. To the best of our knowledge, ECA9 is the first surveillance video dataset for industrial accident scenes.

FireMatch: A semi-supervised video fire detection network based on consistency and distribution alignment
Qinghua Lin, Zuoyong Li, Kun Zeng, Haoyi Fan, Wei Li, Xiaoguang Zhou

Project

• To fully leverage unlabeled data, we extend consistency regularization with a self-adaptive pseudo-label to the video classification field. Generating enough high-quality pseudo-label data for training helps the model achieve accurate fire video classification.
• For addressing the problem of imbalanced labeled and unlabeled data leading to mismatched sampling experiences, we propose video cross-set sample augmentation combined with adversarial distribution alignment to generate additional labeled samples and alleviate this bias.
• We conduct extensive experiments and ablation studies on public datasets and compare our method with state-of-the-art semi supervised methods. The experimental results demonstrate the effectiveness of the proposed method.

Leukocyte classification using relative-relationship-guided contrastive learning
Zuoyong Li, Qinghua Lin, Jiawei Wu, Taotao Lai, Rongteng Wu, David Zhang

Project

• We introduce the heuristic-guided strategy of contrastive learning, which provides positive and negative samples based on relative distance knowledge. This strategy addresses the bottleneck of contrast views in contrast learning resulting from different classes being sensitive to different data expansions.
• We propose a Relative-Relationship-Guided Contrastive Learning Representation (ReCLR) framework for the leukocyte classification, which introduces the prior distance knowledge to mine positive pairs with the adversarial relative relationship and negative pairs with entropy constraint.
• We conduct extensive experiments and compare ReCLR with several state-of-the-art methods on real leukocyte datasets. The results show that our method achieves better classification performance in different evaluation protocols, including linear evaluation, domain transfer, and finetuning, which shows the effectiveness of proposed method.

WtNGAN: Unpaired image translation from white light images to narrow-band images
Qinghua Lin, Zuoyong Li, Kun Zeng, Jie Wen, Yuting Zhang, Jian Chen

Project

• We utilize contrastive learning and structural consistency constraints to facilitate detailed translation from white light images to narrow-band light images while preserving the structure of the generated images.
• For the first time, we introduce Vision Mamba to the field of unpaired image translation. The generator based on Vision Mamba enhances the detailed representation of gastroscopic white light images to narrow-band light images by establishing long-range dependencies.
• We conduct extensive experiments on a self-built gastroscopic dataset and the BraTS2021 dataset, and the experimental results demonstrate that the proposed method outperforms the current state-of-the-art unpaired image translation methods.

DeepCrackAT: An effective crack segmentation framework based on learning multi-scale crack features
Qinghua Lin, Wei Li, Xiangpan Zheng, Haoyi Fan, Zuoyong Li

Project

• We propose a deep crack convolutional network, named DeepCrackAT, which utilizes an Attention mechanism to fuse multi-scale features from convolution and Tokenized MLP for crack segmentation.
• For irregularly distributed cracks, we use the hybrid dilated convolutions to increase the receptive field of convolutional operations and capture more crack features. Additionally, the proposed method employs a tokenized multilayer perceptron to project high-dimensional crack features into a low dimension space, enhancing the network’s ability of noise resistance.
• We introduce the convolutional block attention module to construct an attentional skip-layer fusion block for multi-scale feature fusion. This helps to enhance the network’s perception of the critical crack region and alleviate the problem of information loss in thick crack segmentation.