Feature Extraction Ability Research Articles

MCW: A Generalizable Deepfake Detection Method for Few-Shot Learning

With the development of deepfake technology, deepfake detection has received widespread attention. Although some deepfake forensics techniques have been proposed, they are still very difficult to implement in real-world scenarios. This is due to the differences in different deepfake technologies and the compression or editing of videos during the propagation process. Considering the issue of sample imbalance with few-shot scenarios in deepfake detection, we propose a multi-feature channel domain-weighted framework based on meta-learning (MCW). In order to obtain outstanding detection performance of a cross-database, the proposed framework improves a meta-learning network in two ways: it enhances the model’s feature extraction ability for detecting targets by combining the RGB domain and frequency domain information of the image and enhances the model’s generalization ability for detecting targets by assigning meta weights to channels on the feature map. The proposed MCW framework solves the problems of poor detection performance and insufficient data compression resistance of the algorithm for samples generated by unknown algorithms. The experiment was set in a zero-shot scenario and few-shot scenario, simulating the deepfake detection environment in real situations. We selected nine detection algorithms as comparative algorithms. The experimental results show that the MCW framework outperforms other algorithms in cross-algorithm detection and cross-dataset detection. The MCW framework demonstrates its ability to generalize and resist compression with low-quality training images and across different generation algorithm scenarios, and it has better fine-tuning potential in few-shot learning scenarios.

LGP-YOLO: an efficient convolutional neural network for surface defect detection of light guide plate

Light guide plate (LGP) is a key component of liquid crystal display (LCD) display systems, so its quality directly affects the display effect of LCD. However, LGPs have complex background texture, low contrast, varying defect size and numerous defect types, which makes realizing efficient and accuracy-satisfactory surface defect automatic detection of LGPS still a big challenge. Therefore, combining its optical properties, dot distribution, defect imaging characteristics and detection requirements, a surface defect detection algorithm based on LGP-YOLO for practical industrial applications is proposed in this paper. To enhance the feature extraction ability of the network without dimensionality reduction, expand the effective receptive field and reduce the interference of invalid targets, we built the receptive field module (RFM) by combining the effective channel attention network (ECA-Net) and reviewing large kernel design in CNNs (RepLKNet). For the purpose of optimizing the performance of the network in downstream tasks, enhance the network's expression ability and improve the network’s ability of detecting multi-scale targets, we construct the small detection module (SDM) by combining space-to-depth non-strided convolution (SPDConv) and omini-dimensional dynamic convolution (ODConv). Finally, an LGP defect dataset is constructed using a set of images collected from industrial sites, and a multi-round experiment is carried out to test the proposed method on the LGP detect dataset. The experimental results show that the proposed LGP-YOLO network can achieve high performance, with mAP and F1-score reaching 99.08% and 97.45% respectively, and inference speed reaching 81.15 FPS. This demonstrates that LGP-YOLO can strike a good balance between detection accuracy and inference speed, capable of meeting the requirements of high-precision and high-efficiency LGP defect detection in LGP manufacturing factories.

A Siamese Multiscale Attention Decoding Network for Building Change Detection on High-Resolution Remote Sensing Images

The objective of building change detection (BCD) is to discern alterations in building surfaces using bitemporal images. The superior performance and robustness of various contemporary models suggest that rapid development of BCD in the deep learning age is being witnessed. However, challenges abound, particularly due to the diverse nature of targets in urban settings, intricate city backgrounds, and the presence of obstructions, such as trees and shadows, when using very high-resolution (VHR) remote sensing images. To overcome the shortcomings of information loss and lack of feature extraction ability, this paper introduces a Siamese Multiscale Attention Decoding Network (SMADNet). This network employs the Multiscale Context Feature Fusion Module (MCFFM) to amalgamate contextual information drawn from multiscale target, weakening the heterogeneity between raw image features and difference features. Additionally, our method integrates a Dual Contextual Attention Decoding Module (CADM) to identify spatial and channel relations amongst features. For enhanced accuracy, a Deep Supervision (DS) strategy is deployed to enhance the ability to extract more features for middle layers. Comprehensive experiments on three benchmark datasets, i.e., GDSCD, LEVIR-CD, and HRCUS-CD, establish the superiority of SMADNet over seven other state-of-the-art (SOTA) algorithms.

EfficientUNet: An efficient solution for breast tumour segmentation in ultrasound images

AbstractThe accurate segmentation of breast tumours is important for the diagnosis and treatment of breast cancer. When using the classic U‐Net, Attention‐UNet, and UNet++ to segment the tumour, there are problems of oversegmentation, incorrect segmentation, and poor edge continuity. In this paper, an effective tumour segmentation method, EfficientUNet, is proposed. EfficientUNet adopts a step‐by‐step enhancement method, combining ResNet18, a channel attention mechanism and deep supervision. ResNet18, as the encoder of the whole network, solves the problem of gradient disappearanceand improves the feature extraction ability of the model. The channel attention module makes the model more accurate in tumour edge processing. The deep supervision technology accelerates the model training and provides the convergence direction for the model. In addition, it is found that when adjusting the size of the image, the method of image filling before clipping (or zooming) is more conducive to model learning than the direct interpolation method. And a comparative experiment wasperformed on dataset B. Compared to U‐Net, Attention‐UNet and UNet++, EfficientUNet has the highest performance. Finally, the ablation experiment also indicated the effectiveness of each module in EfficientUNet.

The study and analysis related to the evolution of the knowledge tracing research

In the field of education, the level of awareness of students' knowledge status largely affects students' learning efficiency. Knowledge Tracing, as a method of modeling students' learning status, obtaining their knowledge status to predict their future learning can provide more targeted learning solutions, which is extremely useful for improving students' learning efficiency and implementing tailor-made education. This paper introduces the related research methods and their core ideas in the order of probability-based models, logistic function-based models, and deep learning-based models by investigating the development of this field since its introduction. Among the probability-based models, the mainstream methods include BKT and DBKT, which are based on Hidden Markov Models with certain assumptions and are not ideal in practical application scenarios. In contrast, logistic-based models consider a variety of factors in the student learning process and have achieved good results. After that, deep learning-based model has achieved excellent assessment results through its powerful feature extraction ability, and many variants have been derived. This paper introduces DKT in detail and briefly introduces its variant models. At last, this paper summarizes some of the mainstream evaluation criteria in this field and gives the widely used datasets and their categories. Finally, this paper proposes some suggestions for the future development of the field in view of the shortcomings of the current research status for reference.

A Real-Time Flame Detection Method Using Deformable Object Detection and Time Sequence Analysis.

Timely and accurate flame detection is a very important and practical technology for preventing the occurrence of fire accidents effectively. However, the current methods of flame detection are still faced with many challenges in video surveillance scenarios due to issues such as varying flame shapes, imbalanced samples, and interference from flame-like objects. In this work, a real-time flame detection method based on deformable object detection and time sequence analysis is proposed to address these issues. Firstly, based on the existing single-stage object detection network YOLOv5s, the network structure is improved by introducing deformable convolution to enhance the feature extraction ability for irregularly shaped flames. Secondly, the loss function is improved by using Focal Loss as the classification loss function to solve the problems of the imbalance of positive (flames) and negative (background) samples, as well as the imbalance of easy and hard samples, and by using EIOU Loss as the regression loss function to solve the problems of a slow convergence speed and inaccurate regression position in network training. Finally, a time sequence analysis strategy is adopted to comprehensively analyze the flame detection results of the current frame and historical frames in the surveillance video, alleviating false alarms caused by flame shape changes, flame occlusion, and flame-like interference. The experimental results indicate that the average precision (AP) and the F-Measure index of flame detection using the proposed method reach 93.0% and 89.6%, respectively, both of which are superior to the compared methods, and the detection speed is 24-26 FPS, meeting the real-time requirements of video flame detection.

Phase space reconstruction and self-attention convolutional neural network for sound source ranging with trial data

To improve the feature extraction ability of multipath signals and reduce the dependence on environmental information, this paper studies a method that combines multipath signal feature information mining and neural network in high-dimensional phase space instead of using the received complex acoustic pressure as the input of the neural network directly. Based on the phase space reconstruction (PSR) technology, we propose a hybrid model of the self-attention-based Transformer and convolutional neural network (CNN), named PSR-TransCNN. The covariance matrix of the reconstructed data is constructed as the input of the sparse auto-encoder (SAE). A transfer learning strategy is employed to obtain the optimal weight and offset value of SAE, which serve as the initial values of the PSR-TransCNN. Real-world experimental results demonstrate that the PSR-TransCNN significantly reduces the relative root mean square error (RMSE) by 0.006 compared with the matched field processing (MFP) method and exhibits outstanding generalization and practical application potential compared to other state-of-the-art (SOTA) methods. The PSR-TransCNN achieves an RMSE of 0.046 and an accuracy of 88.75%, successfully classifying sound sources with unknown labels. The proposed method effectively reduces the multipath effects and addresses the mismatch issue in sound source ranging (SSR) within an unexplored ocean environment.

A spatial design layout optimization model based on deep learning in the context of rural revitalization

In recent years, with the transformation of China's economy and social development, solving rural problems and achieving rural revitalization has become an urgent task. The development of information technology, specifically the advancements in deep learning technology, has provided new solutions for rural revitalization. This article aims to optimize the spatial structure of rural residential areas and investigate models for optimizing spatial design layouts within the context of rural revitalization utilizing deep learning techniques, specifically convolutional neural networks (DeepLab v3+). This article conducts in-depth research on the DeepLab v3+model and finds that it has good feature extraction ability in processing spatial information such as rural residential houses and ancillary land. By introducing an attention mechanism in the encoding region, we improved the feature extraction accuracy of the model. On the ImageNet and CoCo2017 datasets, the accuracy indicators PA, IoU, and F1 of the DeepLab v3+model performed well, reaching 98.25 %, 75.61 %, and 86.49 %, as well as 98.31 %, 74.91 %, and 85.49 %, respectively. Furthermore, it was discovered that as the training frequency approached 20, both ResNet50 and DeepLab v3-SE exhibited a generalization accuracy of 86.83 % and 89.26 %, correspondingly. This shows the effectiveness of our model in acquiring and utilizing knowledge during the training. The study further investigated the performance of the model in processing images with noise. When the noise standard deviation is 25, the mean SSIM of the O-DeepLab v3+network reaches 0.943, which is 0.53 % and 1.95 % higher than the DeepLab v3-SE network and traditional DeepLab v3+network, respectively. When the noise standard deviation is 50, the SSIM mean of the O-DeepLab v3+network reaches 0.82, which is 0.61 % and 1.49 % higher than the DeepLab v3-SE network and traditional DeepLab v3+network, respectively. This suggests that our model exhibits robustness when processing noisy images. In summary, the spatial design layout optimization model based on deep learning in the context of rural revitalization studied in this article has high accuracy and stability in processing spatial information such as rural residential houses and ancillary land. This is of great significance for optimizing rural land use, promoting urban-rural land allocation, and achieving rural revitalization. Therefore, this study has certain application value.

Extraction of building from remote sensing imagery base on multi-attention L-CAFSFM and MFFM

Building extraction from high-resolution remote sensing images is widely used in urban planning, land resource management, and other fields. However, the significant differences between categories in high-resolution images and the impact of imaging, such as atmospheric interference and lighting changes, make it difficult for high-resolution images to identify buildings. Therefore, detecting buildings from high-resolution remote sensing images is still challenging. In order to improve the accuracy of building extraction in high-resolution images, this paper proposes a building extraction method combining a bidirectional feature pyramid, location-channel attention feature serial fusion module (L-CAFSFM), and meticulous feature fusion module (MFFM). Firstly, richer and finer building features are extracted using the ResNeXt101 network and deformable convolution. L-CAFSFM combines feature maps from two adjacent levels and iteratively calculates them from high to low level, and from low to high level, to enhance the model’s feature extraction ability at different scales and levels. Then, MFFM fuses the outputs from the two directions to obtain building features with different orientations and semantics. Finally, a dense conditional random field (Dense CRF) improves the correlation between pixels in the output map. Our method’s precision, F-score, Recall, and IoU(Intersection over Union) on WHU Building datasets are 95.17%、94.83%、94.51% and 90.18%. Experimental results demonstrate that our proposed method has a more accurate effect in extracting building features from high-resolution image.

A shale gas production prediction model based on masked convolutional neural network

Shale gas production prediction is of great significance for shale gas exploration and development, as it can optimize exploration strategies and guide adjustments to production parameters for both new and existing wells. However, the dynamic production characteristics of shale gas wells under the influence of multiple factors such as reservoirs, engineering, and production, exhibit complex nonlinear and non-stationary features, leading to low accuracy in predicting shale gas production. To address this issue, a novel masked convolutional neural network (M-CNN) based on masked autoencoders (MAE) is proposed for shale gas production prediction. First, high-dimensional shale gas production data are transformed into images with unknown information using an encoding structure, thereby converting the regression task into images generation task. Then, convolutional neural network is used for image restoration prediction, and the corresponding numerical values at the image positions are extracted as shale gas production prediction results. Specifically, dilated convolution and multi-scale residual structure (MSRS) are developed to improve the feature representation capability of the network. Meanwhile, convolutional block attention module (CBAM) is adopted to enhance the feature extraction ability of the M-CNN. The performance of our method is validated experimentally on shale gas production data of Changning (CN) block in China. The average RMSE, MRE, and R2 on the test sets are 0.211 (104m3/d), 10.9%, and 0.906, respectively, which is much lower than the traditional time series models. Experimental results demonstrate the effectiveness and superiority of the proposed M-CNN method for shale gas production prediction.

Co-learning-assisted progressive dense fusion network for cardiovascular disease detection using ECG and PCG signals

Electrocardiograms (ECGs) and phonocardiograms (PCGs) are two modalities to provide complementary diagnostic information for improving the early detection accuracy of cardiovascular diseases (CVDs). Existing multi-modality methods mainly used the early or late feature fusion strategy, which did not simultaneously utilize the complementary information contained in low-level detail features and high-level semantic features of different modalities. Meanwhile, they were specially designed for the multi-modality scenario with both ECGs and PCGs, without considering the missing-modality scenarios with only ECGs or PCGs in clinical practice. To address these challenges, we developed a Co-learning-assisted Progressive Dense fusion network (CPDNet) for end-to-end CVD detection, with a three-branch interweaving architecture consisting of ECG and PCG modality-specific encoders and a progressive dense fusion encoder, which could be used for both multi-modality and missing-modality scenarios. Specifically, we designed a novel progressive dense fusion strategy, which not only progressively fused multi-level complementary information of different modalities from low-level details to high-level semantics, but also employed the dense fusion during feature fusion at each level to further enrich available multi-modality information through mutual guidance of features at different levels. Meanwhile, the strategy integrated cross-modality region-aware and multi-scale feature optimization modules to fully evaluate the contributions of different modalities and signal regions and enhance the feature extraction ability of the network for multi-scale target regions. Moreover, we designed a novel co-learning strategy to guide the learning process of the CPDNet by combining intra-modality and joint losses, which made each encoder well-trained. This strategy could not only assist our fusion strategy by making modality-specific encoders provide sufficiently discriminative features for the fusion encoder, but also enable the CPDNet to robustly handle missing-modality scenarios by independently using the corresponding modality-specific encoder. Experimental results on public and private datasets demonstrated that our method not only outperformed state-of-the-art multi-modality methods by at least 5.05% for average accuracy in the multi-modality scenario, but also achieved better performance than single-modality models in the missing-modality scenarios.

Wind Turbine Surface Defect Detection Method Based on YOLOv5s-L

In order to solve the problems of low efficiency, time consumption and high costs in the detection of defects on wind turbine surfaces in industrial scenarios, an improved YOLOv5 algorithm for wind turbine surface defect detection is proposed, named YOLOv5s-L. Firstly, the C3 module of YOLOv5s is replaced with the C2f module, which is more abundant in gradient flow, to enhance the ability of feature extraction and feature fusion. Secondly, the Squeeze and Excitation (SE) module is embedded in the YOLOv5 Backbone network to filter out redundant feature information and retain important feature information. Thirdly, the weighted Bidirectional Feature Pyramid Network (BiFPN) is introduced to replace the FPN + PAN, which can achieve a higher level of feature fusion while keeping the weight light. Finally, the Focal Loss function is used to replace the CIOU Loss function of the YOLOv5 algorithm to optimize the training model and improve the accuracy of the algorithm. The experimental results show that, compared with the traditional YOLOv5 algorithm, the average precision mAP is improved by 1.9%, and the frame rate FPS can reach 145 F/s without increasing the model parameters; it can satisfy the requirements for real-time, accurate detection on mobile devices. This method provides effective support for surface defect detection of wind turbines and provides reference for intelligent wind farm operation and maintenance.

Simultaneous denoising and reconstruction of distributed acoustic sensing seismic data via a multicascade deep-learning method

Distributed acoustic sensing (DAS) has been moderately used in seismic exploration as a consequence of its advantages, such as low cost, high resolution, high sensitivity, and strong resistance to high temperatures. Nonetheless, seismic data recorded by the DAS system often are seriously contaminated by a massive amount of background noise and recording gaps, which also decrease the quality of DAS seismic data. Building on its success in numerous fields of data processing, convolutional neural networks (CNNs) also have been applied to seismic data denoising as well as reconstruction and achieved good performance compared with traditional methods. However, relatively few CNN-based studies have been done on the simultaneous denoising and reconstruction of seismic data, especially for the complex DAS seismic data with a low signal-to-noise ratio. In this study, we have developed a multicascade network structure in combination with a channel attention mechanism (MCA-Net) to achieve the simultaneous denoising and reconstruction of DAS seismic data by only using a unified trained model. Particularly, our MCA-Net consists of three branches of cascade networks that can extract abundant features from DAS seismic data with different resolutions. Furthermore, we deploy some channel attention modules and aggressive blocks to MCA-Net to fuse these aforementioned multiresolution features and thus improving the feature extraction ability of the whole network. Meanwhile, a high-quality training data set composed of paired clean complete and noisy incomplete patches is used to train MCA-Net in a supervised fashion. Some synthetic and real DAS seismic data are used to demonstrate the effectiveness of MCA-Net, and our method exhibits better denoising and reconstruction performance than some traditional methods and two existing CNN architectures.

A dual-branch siamese spatial-spectral transformer attention network for Hyperspectral Image Change Detection

The convolutional neural networks have recently gained widespread attention in Hyperspectral Image Change Detection (HSI-CD) due to their outstanding feature extraction ability. However, limited by the inherent network backbones, the convolutional neural networks (CNNs) fail to mine the sequence attributes and model the intricate relationships of spectral signatures. In contrast, transformers are proficient at learning sequence information owing to the powerful self-attention mechanisms. The two backbone structures exhibit complementary spatial and spectral feature extraction strengths, respectively. Inspired by this, we propose a dual-branch siamese spatial–spectral transformer attention network (DBS3TAN) for HSI-CD. The main idea is to fully exploit the advantages of CNNs and transformers for spatial and spectral feature extraction. More importantly, we devise the two key modules, i.e., the spatial attention module and the spatial–spectral transformer module. The former utilizes depthwise separable convolutions and attention mechanisms to emphasize the features of dual-temporal HSIs from the spatial perspective. The latter focuses on the sequence attributes of spectral signatures and mines the spatial characteristics from adjacent pixels. We employ the weighted contrastive loss function to separate the changed and unchanged pixels more reliably and set the random weight factors to balance the contributions of the two branches. Finally, the threshold values judgment is used to obtain the ultimate detection maps. We conduct extensive experiments to evaluate the DBS3TAN on three HSI datasets, demonstrating its superior performances than compared methods qualitatively and quantitatively. The source code will be available at https://github.com/zhangyiyan001/DBS3TAN.

Intelligent prediction for support resistance in working faces of coal mine: A research based on deep spatiotemporal sequence models

Accurately predicting support resistance in coal mine working faces is a pressing matter in the field of intelligent mining research, which holds immense importance in enhancing the safety of coal production and the overall level of intelligent mining. The current research in this field primarily focuses on utilizing deep temporal sequence models for prediction. While there have been promising advancements in this area, concerns persist regarding the dependability of predicted outcomes. In fact, the movement of overburden in mining faces varies over time and space, and the changes in the support resistance that reflect the instability of overburden movement also have temporal and spatial characteristics. Therefore, in order to improve the reliability of model prediction, this paper proposes the application of deep spatiotemporal sequence models for support resistance prediction, and identifies the PredRNN and PredRNN++ models as suitable options. The exploration on the application of deep spatiotemporal sequence models for predicting the support resistance has been conducted. The proposed application involves the establishment of a fully working face area grid model with support number and mining cycle as coordinate units. Different from deep temporal sequence models that take the studied grid features as observations, the 25 adjacent spatiotemporal feature grids corresponding to the studied grid are considered as observations for the deep spatiotemporal sequence models. Based on the predicted three-dimensional tensor, the required predicted support resistance data can be obtained. Based on the complete support resistance data from 9 working faces in a coal mine in Datong mining area from 2009 to 2018, a comprehensive dataset was established to experimentally analyze the predictive performance and reliability of the LSTM, PredRNN, and PredRNN++ models. The experimental results demonstrate significant improvements in predictive performance and reliability of deep spatiotemporal sequence models compared to the current mainstream deep temporal sequence models. In terms of predictive performance, the PredRNN and PredRNN++ models have improved their feature extraction ability by incorporating the convolution operations to extract the spatial correlation of the input data. Furthermore, through the incorporation of gated units and nonlinear neurons into the storage units, coupled with the integration of vertical memory state transitions, the depth of the network has significantly increased, resulting in the enhancement of the sequence modeling capabilities. In terms of reliability of prediction results: The acceptable threshold for prediction error was set at 1000 kN. According to the average prediction results of all samples in the test set, it was found that the deep temporal sequence model was unable to accurately predict support resistance ahead. While the deep spatiotemporal sequence model demonstrated a predictable sequence length of 8. According to the prediction results of all samples in the test set, the accurate prediction rate increased from 30 %∼35 % to 65 %∼85 % and the maximum MAE value decreased from 10000 ∼ 12500kN to 6500 ∼ 9000kN in more than 20,000 test samples. The PredRNN model with an input sequence length of 50 achieves an accurate prediction rate of 85.48 % and a maximum MAE value of 6698kN, making it the most reliable prediction model.

YOLOv5-CD: Strip steel surface defect detection method based on coordinate attention and a decoupled head

The ability to detect strip steel surface defects is crucial for ensuring the quality of items made from that material. A YOLOv5-CD algorithm is proposed to overcome the problems of insufficient feature extraction ability, low detection accuracy, and slow convergence rate of the strip steel surface defect detection technique in an industrial scenario. First, a coordination attention mechanism is added to the backbone network. This strategy embeds positional information into channel attention, addressing the issue of positional information loss caused by global pooling and effectively improving the model's feature extraction capacity. Second, the head component uses a decoupled head detector. By separating the classification and regression tasks, this approach effectively increases the model's detection precision and convergence rate. According to the experimental findings, the mAP of the YOLOv5-CD algorithm on the NEU-DET dataset is 80.6 %, 5.8 % higher than that of the original YOLOv5 method. Additionally, the algorithm's convergence speed is faster and better able to meet the demands of actual industrial production.

A Forest Wildlife Detection Algorithm Based on Improved YOLOv5s.

A forest wildlife detection algorithm based on an improved YOLOv5s network model is proposed to advance forest wildlife monitoring and improve detection accuracy in complex forest environments. This research utilizes a data set from the Hunan Hupingshan National Nature Reserve in China, to which data augmentation and expansion methods are applied to extensively train the proposed model. To enhance the feature extraction ability of the proposed model, a weighted channel stitching method based on channel attention is introduced. The Swin Transformer module is combined with a CNN network to add a Self-Attention mechanism, thus improving the perceptual field for feature extraction. Furthermore, a new loss function (DIOU_Loss) and an adaptive class suppression loss (L_BCE) are adopted to accelerate the model's convergence speed, reduce false detections in confusing categories, and increase its accuracy. When comparing our improved algorithm with the original YOLOv5s network model under the same experimental conditions and data set, significant improvements are observed, in particular, the mean average precision (mAP) is increased from 72.6% to 89.4%, comprising an accuracy improvement of 16.8%. Our improved algorithm also outperforms popular target detection algorithms, including YOLOv5s, YOLOv3, RetinaNet, and Faster-RCNN. Our proposed improvement measures can well address the challenges posed by the low contrast between background and targets, as well as occlusion and overlap, in forest wildlife images captured by trap cameras. These measures provide practical solutions for enhanced forest wildlife protection and facilitate efficient data acquisition.

Underwater Image Translation via Multi-Scale Generative Adversarial Network

The role that underwater image translation plays assists in generating rare images for marine applications. However, such translation tasks are still challenging due to data lacking, insufficient feature extraction ability, and the loss of content details. To address these issues, we propose a novel multi-scale image translation model based on style-independent discriminators and attention modules (SID-AM-MSITM), which learns the mapping relationship between two unpaired images for translation. We introduce Convolution Block Attention Modules (CBAM) to the generators and discriminators of SID-AM-MSITM to improve its feature extraction ability. Moreover, we construct style-independent discriminators that enable the discriminant results of SID-AM-MSITM to be not affected by the style of images and retain content details. Through ablation experiments and comparative experiments, we demonstrate that attention modules and style-independent discriminators are introduced reasonably and SID-AM-MSITM performs better than multiple baseline methods.

Few-Shot Air Object Detection Network

Focusing on the problem of low detection precision caused by the few-shot and multi-scale characteristics of air objects, we propose a few-shot air object detection network (FADNet). We first use a transformer as the backbone network of the model and then build a multi-scale attention mechanism (MAM) to deeply fuse the W- and H-dimension features extracted from the channel dimension and the local and global features extracted from the spatial dimension with the object features to improve the network’s performance when detecting air objects. Second, the neck network is innovated based on the path aggregation network (PANet), resulting in an improved path aggregation network (IPANet). Our proposed network reduces the information lost during feature transfer by introducing a jump connection, utilizes sparse connection convolution, strengthens feature extraction abilities at all scales, and improves the discriminative properties of air object features at all scales. Finally, we propose a multi-scale regional proposal network (MRPN) that can establish multiple RPNs based on the scale types of the output features, utilizing adaptive convolutions to effectively extract object features at each scale and enhancing the ability to process multi-scale information. The experimental results showed that our proposed method exhibits good performance and generalization, especially in the 1-, 2-, 3-, 5-, and 10-shot experiments, with average accuracies of 33.2%, 36.8%, 43.3%, 47.2%, and 60.4%, respectively. The FADNet solves the problems posed by the few-shot characteristics and multi-scale characteristics of air objects, as well as improving the detection capabilities of the air object detection model.

MFSPest: A multi-scale feature selection network for light-trapped agricultural pest detection

Accurate and fast pest detection is crucial for ensuring high crop yield and quality in modern agriculture. However, there are significant challenges in using deep learning for pest detection, such as the small proportion of pest individuals in the image area, complex backgrounds in light-trapped pest images, and an unbalanced distribution of pest species. To address these problems, we propose MFSPest, a multi-scale feature selection network for detecting agricultural pests in trapping scenes. We design a novel selective kernel spatial pyramid pooling structure (SKSPP) in the feature extraction stage to enhance the network’s feature extraction ability for key regions and reduce its focus on irrelevant background information. Furthermore, we present the equalized loss to increase the loss weights of rare categories and improve the distribution imbalance among pest categories. Finally, we build LAPD, a light-trapping agricultural pest dataset containing nine pest categories. We conducted experiments on this dataset and demonstrated that our proposed method achieves state-of-the-art performance, with Accuracy, Recall, and mean Average Precision (mAP) of 89.9%, 92.8%, and 93.6%, respectively. Our method satisfies the requirements of pest detection applications in practical scenarios and has practical value and economic benefits for use in agricultural pest trapping and management.