To achieve the non-destructive detection of peanut defects, a multi-target identification method based on the multispectral system and improved Faster RCNN is proposed in this paper. In terms of the system, the root-mean-square contrast method was employed to select the characteristic wavelengths for defects, such as mildew spots, mechanical damage, and the germ of peanuts. Then, a multispectral light source system based on a symmetric integrating sphere was designed with 2% nonuniformity illumination. In terms of Faster RCNN improvement, a texture-based attention and a feature enhancement module were designed to enhance the performance of its backbone. In the experiments, a peanut-deficient multispectral dataset with 1300 sets was collected to verify the detection performance. The results show that the evaluation metrics of all improved compared with the original network, especially in the VGG16 backbone network, where the mean average precision (mAP) reached 99.97%. In addition, the ablation experiments also verify the effectiveness of the proposed texture module and texture enhancement module in peanut defects detection. In conclusion, texture imaging enhancement and efficient extraction are effective methods to improve the network performance for multi-target peanut defect detection.

Hyperspectral image (HSI) classification has become a popular research topic in recent years, and transformer-based networks have demonstrated superior performance by analyzing global semantic features. However, using transformers for pixel-level HSI classification has two limitations: ineffective capture of spatial-spectral correlations and inadequate exploitation of local features. To address these challenges, we propose a Dual-Dimension Self-Attention (D ² SA) mechanism that fully exploits HSI's high spectral-spatial correlation by using two separate branches to model the global dependence of features from the spectral and spatial dimensions. Additionally, we develop a multi-layer residual convolution module that extracts local features and introduces shallow-deep feature interactions to obtain more discriminative representations. Based on these components, we propose a Dual Dimension Spectral-Spatial Bottleneck Transformer (D ² S ² BoT) framework for HSI classification that simultaneously models the local interactions and global dependencies of HSI pixels to achieve high-precision classification. By virtue of the D ² SA mechanism, the introduced D ² S ² BOT framework can produce competitive classification results with a limited number of training samples on three well-known datasets, which we hope will provide a strong baseline for future research on transformers in the field of HSI. Codes are released at: https://github.com/HyperSystemAndImageProc/D2S2BoT .

To exploit hyperspectral image's (HSI) spectral–spatial information and reduce network complexity, a triple-path convolution neural network with an interleave-attention mechanism is constructed for high-precision classification. A hybrid branch is proposed to capture joint features, which are later utilized as complementary information for purely spectral and spatial features. Furthermore, the interleave-attention mechanism is elaborately designed to increase the interaction of data from spectral, spatial, and joint branches as they propagate through the network for feature integration. Meanwhile, two attention modules are adopted in the corresponding branch to optimize extracted features for better feature representation. We utilize several real HSI datasets to evaluate network performance, which demonstrates that the proposed triple-path network can obtain very satisfactory performance with fewer parameters and low computational complexity.

Deep learning-based methods for hyperspectral image super-resolution (HISR) have recently achieved remarkable success. However, the contextual correlation and intrinsic properties of adjacent bands in hyperspectral images have not been fully exploited in previous works. Therefore, this paper proposes a network framework (TC-HISRNet) that leverages the local feature extraction capability of convolutional neural networks (CNN) in tandem with the global feature extraction ability of transformer, to thoroughly investigate the role of contextual information in HISR. TC-HISRNet is comprised of a two-stage feature encoding network, which combines a two-channel convolutional neural network (CNN) and a transformer encoder. Within the CNN encoder, the joint utilization of the real-time attention fusion module and 2D/3D convolutions enables the preservation of spectral correlation between bands while facilitating the extraction of local features from hyperspectral images. On the other hand, the developed transformer encoder leverages joint adjacent band contextual information to capture long-range dependencies and interdependencies among the combined features. Extensive evaluation and comparison of three public datasets demonstrated that the proposed method outperformed the state-of-the-art super-resolution (SR) method, resulting in superior reconstruction of hyperspectral images.

Accurate wafer surface classification is essential for semiconductor process control, but its practical application is limited by the small number of labeled hyperspectral samples and the continuous emergence of previously unseen surface classes. To address these challenges, this study proposes a diffusion-model-enhanced incremental learning framework for hyperspectral wafer surface classification. A hyperspectral dataset containing eight representative wafer surface classes was constructed, covering normal material or structural states, process-related surface conditions, and potential defects. The diffusion model jointly learns the spatial structures and spectral correlations of hyperspectral data and generates supplementary samples for classifier training. During incremental updates, generated samples from previously learned classes are further used as rehearsal data to preserve historical knowledge, while the classification head is expanded to accommodate newly introduced classes. Experimental results show that diffusion-based augmentation improves the accuracy and Macro F 1 score of the convolutional neural network (CNN) by more than three percentage points compared with the non-augmented setting. In the continual-learning comparison, the proposed method achieves an accuracy of 95.98% and a Macro F 1 score of 95.97% on the diffusion-enhanced dataset. The forgetting rate decreases from 6.79% to 4.64%, corresponding to a relative reduction of 31.66%. These results demonstrate that diffusion-generated rehearsal can improve hyperspectral wafer surface classification while effectively mitigating catastrophic forgetting during incremental updates.