Understanding and Mitigating Over-refusal for Large Language Models via Safety Representation

Junbo Zhang, Ran Chen, Qianli Zhou, Xinyang Deng, Wen Jiang

We analyze over-refusal in LLMs from a representation perspective, finding such samples lie at the boundary of benign and malicious ones. We propose MOSR with Overlap-Aware Loss Weighting and Context-Aware Augmentation to intervene safety representation. Experiments show we mitigate over-refusal while maintaining safety, offering insights for balancing model safety and over-refusal.

Unsupervised global-local domain adaptation with self-training for remote sensing image semantic segmentation

Junbo Zhang, Zhiyong Li, Mantao Wang, Kunhong Li

We propose a hybrid framework integrating global-local adversarial training and self-training. We achieve feature alignment via multi-level discriminators, acquire domain-specific knowledge through self-training, and correct pseudo-labels with a self-labeling mechanism, effectively alleviating global-local domain shift for remote sensing image semantic segmentation.

Unsupervised adversarial domain adaptation for agricultural land extraction of remote sensing images

Junbo Zhang, Shifeng Xu, Jun Sun, Dinghua Ou, Xiaobo Wu, Mantao Wang

We adopt unsupervised adversarial domain adaptation for agricultural land extraction. We use Transformer as the generator backbone with a multi-scale feature fusion module, reduce source-target domain gap via GAN, and minimize intra-domain differences by dividing the target domain with an entropy-based method, enabling accurate unsupervised extraction.

Fusing spatial attention with spectral-channel attention mechanism for hyperspectral image classification via encoder–decoder networks

Jun Sun*, Junbo Zhang*, Xuesong Gao, Mantao Wang, Dinghua Ou, Xiaobo Wu, Dejun Zhang

We propose an encoder-decoder network fusing spatial and spectral-channel attention, with three fusion strategies to utilize dual attention. We use a hierarchical Transformer in the encoder to extract long-range context, and fuse multi-scale features via upsampling and skip connections in the decoder, improving hyperspectral image classification performance.

BAFormer: A Novel Boundary-Aware Compensation UNet-like Transformer for High-Resolution Cropland Extraction

Zhiyong Li, Youming Wang, Fa Tian, Junbo Zhang, Yijie Chen, Kunhong Li

To address inaccurate boundary segmentation in heterogeneous croplands, this paper proposes BAFormer, a UNet-like boundary-aware model. It targets high-frequency boundary features by integrating a Feature Adaptive Mixer (FAM) for adaptive feature acquisition and a Depthwise Large Kernel MLP (DWLK-MLP) to supplement positional information via large receptive fields. Evaluations on datasets like Vaihingen demonstrate superior performance, with the lightweight BAFormer-T notably outperforming existing models.

YOLOv7-Plum: advancing plum fruit detection in natural environments with deep learning

Rong Tang, Yujie Lei, Beisiqi Luo, Junbo Zhang, Jiong Mu

We propose YOLOv7-Plum, an improved YOLOv7-based model for plum fruit detection. We construct a dataset of plums in natural environments, introduce CBAM attention, replace CSPSPP with CSPSPPF, and use bilinear interpolation for upsampling, enhancing detection efficiency and accuracy.

Multi-Scale Spatial Attention-Based Multi-Channel 2D Convolutional Network for Soil Property Prediction

Guolun Feng, Zhiyong Li, Junbo Zhang, Mantao Wang

To address low accuracy in VNIR soil prediction models, we propose a Convolutional Neural Network (CNN) utilizing 2D multi-channel inputs and a multi-scale spatial attention mechanism. By applying the Gramian Angular Field (GAF) method to the LUCAS dataset, this approach significantly enhances feature extraction. Statistical comparisons confirm that the model outperforms current state-of-the-art techniques, achieving high precision across seven key soil properties.