Customs documents contain a large amount of information. The image quality of customs declaration forms is often poor, with issues such as blurred characters, missing strokes, stroke merging, and noise contamination, which increase the difficulty of detection and recognition. Previous approaches for text detection and recognition have already achieved promising performances across various benchmarks. However, they usually fall short when dealing with customs declaration forms. In this paper, we propose EFF-DBNet, an efficient text detector, and SKSVTR, an advanced text recognition model. In the detection stage, we redesign the network based on the DBNet. We introduce the Efficient Channel Attention to extract features from image and utilize Feature Pyramid Enhancement Modules (FPEMs) for feature enhancement, then use a Feature Fusion Module to gather the features given by FPEMs of different depths into a final feature. For the recognition stage, we improve the SVTR model by replacing the local mixing in the backbone network with convolutional modules, and then introduce the Selective Kernel Attention to effectively capture multi-scale features in the image space. Experimental results on customs declaration forms demonstrate the effectiveness of our methods, which outperform existing methods.

In industrial scenarios, existing deep generative methods face the challenge of adapting to new domains with limited data for anomaly detection. Flow-based generative models are efficient but fail to perform well in out-of-distribution detection due to the inductive bias. To solve these problems, we utilize a hierarchical normalizing flow framework for few-shot anomaly detection and localization (FesFlow). First, we propose a multi-scale attention coupling block in each single flow with channel and spatial self-attention mechanism, which can model long-term dependencies of flows and obtain fine-grained anomalous feature distribution. Furthermore, to reduce the bias towards local-pixel correlation, we introduce a global and local aggregation module to capture semantic context and fuse with low-level detailed features. Compared to state-of-the-art methods, our approach demonstrates outstanding performance evaluated on MVTec-AD and BTAD datasets, achieving optimal balances in both detection and localization tasks in an end-to-end network.

Recently, implicit function-based approaches have advanced 3D human reconstruction from a single-view image. However, previous methods suffer from issues such as noisy artifacts, loss of geometric details, and broken limbs under the scenarios of challenging poses. To address these problems, a novel end-to-end deep neural network named ReGIR is proposed, which is a multi-level architecture combining the parametric model with implicit function. The architecture consists of a coarse level and a fine level, and for each level, normal maps and the signed distance function (SDF) are introduced to encode query points. Furthermore, the network is trained in a coarse-to-fine manner to enable robust human body reconstruction with geometric details. Our extensive qualitative and quantitative experiments demonstrate that ReGIR achieves competitive reconstruction results.

Focusing on the challenge of reconstructing 3D model from single-view image of clothed human, we propose a novel method named Periodic Implicit Neural Representation (PeINR) within the framework of PaMIR, which is a method for human reconstruction by combining the parametric human body model with the free-form deep implicit function. This approach addresses several critical aspects to improve the accuracy and realism of reconstructed models. Firstly, we introduce normal maps as a means to incorporate geometric priors, thereby enhancing the precision of the back reasoning process. Secondly, an attention mechanism is integrated to augment the image feature encoder, facilitating the extraction of more comprehensive image features critical for accurate reconstruction. Finally, our method employs an implicit representation based on a sinusoidal periodic function. This allows for the accommodation of more complex information, significantly enhancing the capability of the model to capture fine geometric details. Compared to previous methods, PeINR stands out by not only faithfully reconstructing realistic and natural clothed human body mesh but also by effectively portraying the intricate geometry and texture details in the 3D human model. This innovative approach successfully mitigates issues encountered in prior methods, such as bumps, broken limbs, and unnatural poses during the reconstruction process. Experimental results validate the efficacy of PeINR, showcasing remarkable outcomes in single-view 3D clothed human reconstruction. Our architecture demonstrates superior performance in addressing challenges associated with reconstructing detailed and lifelike human model.

Few-shot object detection aims to identify objects from limited annotated samples. It has received extensive attention in recent years and has made great progress. Attention methods currently applied to the few-shot problem often ignore the interaction between features of different dimensions. This will cause the model uable to pay enough attention to important areas., thus affecting the detection accuracy. In addition, existing methods for obtaining class prototypes usually rely on simple global average pooling, which cannot make the prototype contain perceptual information and is not representative enough, resulting in the generated class prototype not accurately reflecting the characteristics of the target class. To address these problems, our method is based on a two-stage fine-tuning paradigm, studies the dimensional interaction of features, and proposes a Hybrid Attention network to acquire attention weights by obtaining the interaction between different feature dimensions. To make the prototype features contain more representative perceptual information, we propose a Prototype Fusion Network to allow the information of the query image to participate in the generation process of the prototype features. Our method has achieved competitive improvements based on the baseline method, and its performance has been verified by extensive experiments on the general datasets PASCAL VOC and MS COCO. For the general dataset PASCAL VOC, our method improves the AP50 on split 1 by 9.3% on 1-shot and 7.7% on 3-shot. Index Terms─

Various defects arise during textile production, making fabric defect inspection essential for production and quality management in the textile industry. However, fabric defect detection techniques face challenges due to defects with disparate aspect ratios, high foreground-background similarity, and tiny sizes. Based on these issues, we propose a fabric defect detection method that combines an improved Cascade R-CNN (SPCNet) network and Super-Resolution reconstruction technology. Firstly, defective images are reconstructed using a hybrid attention Transformer (HAT), enhancing texture details and edge information. We design a new multi-stage defect detection model SPCNet to identify fabric defects. The architecture includes a feature extraction module based on Switchable Atrous Convolution (SAC). SAC can obtain a larger receptive field for better detection of tiny defects. Path Aggregation Network (PANet) is introduced to improve the recognition of scale-unbalanced defects. In addition, Cascade RPN (C-RPN) is adopted to fully use deep and shallow features. To solve the issue of imbalanced defect classes, we adopt Class-aware Sampling (CAS) strategy. Soft-Nmsis used to reduce the false deletion of defective feature detection boxes. Comparative experimental results demonstrate that the defect detection method combining HAT and SPCNet can significantly raise the overall recognition rate of multiclass fabric defects, exceeding the performance of other current methods.

Abstract Automatic facial expression analysis remains challenging due to its low recognition accuracy and poor robustness. In this study, we utilized active learning and support vector machine (SVM) algorithms to classify facial action units (AU) for human facial expression recognition. Active learning was used to detect the targeted facial expression AUs, while an SVM was utilized to classify different AUs and ultimately map them to their corresponding facial expressions. Active learning reduces the number of non-support vectors in the training sample set and shortens the labeling and training times without affecting the performance of the classifier, thereby reducing the cost of labeling samples and improving the training speed. Experimental results show that the proposed algorithm can effectively suppress correlated noise and achieve higher recognition rates than principal component analysis and a human observer on seven different facial expressions.