Though object detection has been widely used in the industrial scene, it still faces the detection problems of crack defects with slender and rotated characteristics. On the one hand, traditional horizontal anchor methods are usually hard to precisely locate the object. On the other hand, CNNs (Convolutional Neural Networks) perform poorly in terms of feature extraction from rotated objects. In addition, normal loss functions are insensitive to slender objects. To address these, this paper proposes a Slender and Rotated Detector (SR-Det) for robust slender and rotated object detection. Specifically, the Rotated Region Calibration (RRC) is designed, which takes horizontal proposals with different scales and aspect ratios as inputs and outputs the corresponding rotation proposals. Then, the Rotated Angle Proposal Align (RAP-Align) is presented to guarantee the quality of extracted feature information. Finally, the Rotated intersection over union(R-IoU) based on Intersection Over Union (IoU) strategy is proposed for guiding the model to maximize the area between predicted box and Ground Truth box. The experiments on metal cans and curtain walls datasets have shown that the method proposed achieves state-of-the-art performance, demonstrating the effectiveness of the proposed algorithm.

How to extract important features by existing metric-based few-shot image classification models is a difficulty. A few-shot image classification method based on local orthogonal feature fusion is proposed. First, the feature extraction network is used to simultaneously extract shallow features with rich local details and deep features with strong semantics. Then, a channel attention module and a multi-scale feature adaptive fusion module are used to perform feature enhancement on the channel and scale dimensions of the shallow features, respectively, in order to generate the feature with more salient local features and more scale information. Finally, according to local orthogonal feature extraction and attention fusion, the obtained multi-scale local features and initial deep semantic features are extracted and fused by a local orthogonal feature fusion module. In this way, we can make full use of the local and global feature information of the image. And a feature representation is generated, which can be more representative of the target category. The experimental results on the three public datasets of miniImageNet, tieredImageNet and CUB-200-2011 show that the proposed method can achieve better classification results. The accuracy rate of the proposed method on the 5way-5shot task reaches 81.69%, 85.36% and 89.78% respectively. Compared with the baseline model, the classification accuracy increased by 5.23%, 3.19% and 5.99% respectively.

In traditional anchor box-based text detection methods for natural scenes, anchor boxes are prone to interference from other text instances, resulting in erroneous judgments or affecting accuracy. Moreover, text instances contain strong topological features, which are usually be ignored, resulting in poor performance in curved circular text detection tasks. To solve this problem, a novel neural network structure is proposed, which introduces the concept of graph convolutional networks by fully considering the relationship between adjacent anchor frames, and incorporating the topological characteristics of anchor frames to assist the learning of graph neural networks, improving the effectiveness of the overall network. The ablation experiments were conducted on two publicly available natural scene text detection datasets. In the CTW1500 dataset, the proposed method improved the model by approximately 3.0%, 1.9%, and 2.5% in terms of recall, accuracy, and F-score, respectively, and in the Totel-Text dataset , the three values were improved by approximately 2.2%, 1.8%, and 2.0%, respectively. In addition, the proposed method has also been compared with other text detection algorithms proposed in recent years. Experimental results show that the proposed method performs well for text detection in complex natural scenes, demonstrating the promising effectiveness of the proposed module for improving the performance of text detection.

Most existing industrial anomaly detection methods adopt the idea of image restoration, which utilizes artificial synthesis of pseudo anomaly samples in the training stage to discriminate the artificial defect image from the reconstructed images after model restoration, and calculates the deviation of the discrimination to obtain the anomalous region. However, the artificial defect images do not correlate well with the actual defect semantics and cannot accurately cover the actual defect types, resulting in less robustness of the model. In order to solve this problem, this paper proposes a two-stage industrial Diffusion Anomaly Detection network (DADNet). Firstly, a semantically-guided anomaly generation module is utilized to synthesize the known defects, and this is used as the a priori information for anomaly detection. In the first stage, an image reconstruction model is used to repair the abnormal region of the workpiece. In the second stage, a discriminative model based on the repaired image and the initial image is trained and used to detect the anomalous regions. In addition, this paper further enhances the performance of the reconstruction model by aggregating features through the joint attention mechanism. DADNet outperforms the existing models on the publicly available datasets of workpieces with different materials, promising prospect for industrial anomaly defect detection.

Automated radiology report generation is crucial for reducing radiologist workload and minimizing diagnostic errors. Although existing studies have conducted in-depth research on lesion regions, there is potential for enhancement in generating detailed descriptions. Current methods tend to diminish sensitivity to the semantic information of visual lesions and weaken the critical association between visual and textual semantics. This paper introduces a novel Prior Prompt-Driven Semantic Consistency Model (PPD-SCM) to address these limitations. The Prompt-Lesion Enhancement module in the proposed model systematically integrates both normal and abnormal diagnostic descriptions from radiological chest X-ray images to construct prior prompts. By employing a prompt attention mechanism that fuses visual features with textual prompts, this module enhances the model's ability to perceive potential lesion features. Furthermore, this study introduces a Visual-Textual Semantic Consistency (VTSC) module that employs contrastive learning to deeply align visual and textual semantics. By leveraging prompt tokens to guide the model in generating enriched contextual information, the VTSC optimizes the subsequent report generation process. It effectively reduces the semantic gap between medical images and the generated reports, thereby enhancing the accuracy and reliability of report generation. Extensive experimental results on the IU X-Ray and MIMIC-MV datasets demonstrate that our proposed method significantly outperforms existing approaches in generating high-quality radiology reports.

Currently, the main approach of Alzheimer's Disease (AD) diagnosis is realized by structural magnetic resonance imaging (sMRI) , and the existing deep learning-based AD diagnosis is mainly based on 2D (converting 3D sMRI to 2D slices) or 3D convolutional neural networks, which cannot effectively capture 3D sMRI global features. To address this, this paper improves Swin Transformer to realize 3D block division for global features extraction and constructs a Transformer predictive classification model. Due to the sensitivity of existing Alzheimer's patients' regions with atrophy to the transformation of dimensions, the existing deep learning models are not capable of localizing the lesion regions. To overcome this, we propose an adaptive feature extraction module (AFEM) to realize the deformable adaptive feature extraction, and extend the basic 3D Transformer model to construct the AFEM-Transformer deep learning model to further enhance the feature learning ability of the model and realize adaptive localization of the specific location of the pathological region, which can be used to assist clinical diagnosis and realize the classification and prediction of Alzheimer's disease. In this study, sMRI images of 2248 subjects provided by the Alzheimer's Disease Neuroimaging Initiative (ADNI) were selected as the experimental dataset. The proposed AFEM-Transformer model for Alzheimer's disease diagnosis and mild cognitive impairment (MCI) progression prediction tasks will be evaluated and compared with existing convolutional neural network-based models and basic Transformer models. The results show that the experimental results of accuracy, sensitivity, specificity, and area under curve (AUC) value of the proposed AFEM-Transformer model on the two tasks show significant performance improvement compared to the convolutional neural network-based models and basic Transformer model, demonstrating the effectiveness of the proposed AFEM module. The proposed AFEM-Transformer deep learning model is able to accurately diagnose Alzheimer's disease and predict the progression of MCI, and can automatically localize the lesion area, which can be used as an effective computer-aided method in the clinical diagnosis of Alzheimer's disease.

Automatic medical report generation technology plays an important role in auxiliary diagnosis and can greatly reduce the workload of medical workers. As deep learning continues to develop in the medical field, automatic medical report generation technology has become one of the research hotspots. Currently, the main challenges in medical report generation are (1) the difficulty of capturing lesion regions in images by models, and (2) the large semantic gap between visual and language semantics, whose consistency problem is still not well solved. Therefore, in order to solve the above problems, a Cross-Modal Discrepancy Attention Network (CDAN) is proposed to bring closer the semantics between different modalities. The network includes a Reverse Attention (RA) module and a Semantic Consistency (SC) module: (1) the Reverse Attention module explores important areas in medical images more comprehensively, and (2) the Semantic Consistency module utilizes the features of the large language model as a reference to guide the visual features to continuously approach the reference language features, so that the visual semantics can be more accurately converted into language semantics. Experiments show that the Cross-Modal Discrepancy Attention Network is better than the previous model on both IU X-Ray and MIMIC-CXR public datasets, with BLEU4 scores reaching 17.9% and 10.9% respectively. Compared with the baseline model, improvement is significant in performance, which proves that the proposed model is capable of generating accurate and fluent medical reports.