Convolutional Block Research Articles

FINE VESSEL SEGMENTATION WITH REFINEMENT GATE IN DEEP LEARNING ARCHITECTURES

Automated vessel segmentation is essential in diagnosing eye-related disorders and monitoring progressive retinal diseases. State-of-the-art methods have achieved excellent results in this field, but very few have considered the post-processing of feature maps. As a result, there is often a lack of small and fine vessels or discontinuities in segmented vessels. To address this issue, this study introduces a novel post-processing technique called the refinement gate, which works with a deep learning model during training. The refinement gate enhances contextual information to extract important features from feature maps better. The proposed technique is applied with U-net architecture and placed after every convolution block in the encoder path. Visual and statistical comparisons demonstrate the robustness of the proposed method using three publicly available datasets, namely: the DRIVE DB, the STARE DB, and CHASE_DB1 datasets, showing significant improvements to segment weak and tiny vessels. The reported results confirm the potential of the model to be used as a segmentation tool in the medical field. This study is the first to propose such a gating mechanism without additional trainable parameters or standalone networks as in other literature.

RMCNet: A Liver Cancer Segmentation Network Based on 3D Multi-Scale Convolution, Attention, and Residual Path

Abdominal CT images are important clues for diagnosing liver cancer lesions. However, liver cancer presents challenges such as significant differences in tumor size, shape, and location, which can affect segmentation accuracy. To address these challenges, we propose an end-to-end 3D segmentation algorithm, RMCNet. In the shallow encoding part of RMCNet, we incorporated a 3D multiscale convolution (3D-Multiscale Convolution) module to more effectively extract tumors of varying sizes. Moreover, the convolutional block attention module (CBAM) is used in the encoding part to help the model focus on both the shape and location of tumors. Additionally, a residual path is introduced in each encoding layer to further enrich the extracted feature maps. Our method achieved DSC scores of 76.56% and 72.96%, JCC scores of 75.82% and 71.25%, HD values of 11.07 mm and 17.06 mm, and ASD values of 2.54 mm and 10.51 mm on the MICCAI 2017 Liver Tumor Segmentation public dataset and the 3Dircadb-01 public dataset, respectively. Compared to other methods, RMCNet demonstrates superior segmentation performance and stronger generalization capability.

Implementing Real-Time Image Processing for Radish Disease Detection Using Hybrid Attention Mechanisms

This paper developed a radish disease detection system based on a hybrid attention mechanism, significantly enhancing the precision and real-time performance in identifying disease characteristics. By integrating spatial and channel attentions, this system demonstrated superior performance across numerous metrics, particularly achieving 93% precision and 91% accuracy in detecting radish virus disease, outperforming existing technologies. Additionally, the introduction of the hybrid attention mechanism proved its superiority in ablation experiments, showing higher performance compared to standard self-attention and the convolutional block attention module. The study also introduced a hybrid loss function that combines cross-entropy loss and Dice loss, effectively addressing the issue of class imbalance and further enhancing the detection capability for rare diseases. These experimental results not only validate the effectiveness of the proposed method, but also provide robust technical support for the rapid and accurate detection of radish diseases, demonstrating its vast potential in agricultural applications. Future research will continue to optimize the model structure and computational efficiency to accommodate a broader range of agricultural disease detection needs.

Log Volume Measurement and Counting Based on Improved Cascade Mask R-CNN and Deep SORT

Logs require multiple verifications to ensure accurate volume and quantity measurements. Log end detection is a crucial step in measuring log volume and counting logs. Currently, this task primarily relies on the Mask R-CNN instance segmentation model. However, the Feature Pyramid Network (FPN) in Mask R-CNN may compromise accuracy due to feature redundancy during multi-scale fusion, particularly with small objects. Moreover, counting logs in a single image is challenging due to their large size and stacking. To address the above issues, we propose an improved log segmentation model based on Cascade Mask R-CNN. This method uses ResNet for multi-scale feature extraction and integrates a hierarchical Convolutional Block Attention Module (CBAM) to refine feature weights and enhance object emphasis. Then, a Region Proposal Network (RPN) is employed to generate log segmentation proposals. Finally, combined with Deep SORT, the model tracks log ends in video streams and counts the number of logs in the stack. Experiments demonstrate the effectiveness of our method, achieving an average precision (AP) of 82.3, APs of 75.3 for small, APm of 70.9 for medium, and APl of 86.2 for large objects. These results represent improvements of 1.8%, 3.7%, 2.6%, and 1.4% over Mask R-CNN, respectively. The detection rate reached 98.6%, with a counting accuracy of 95%. Compared to manually measured volumes, our method shows a low error rate of 4.07%.

A deep transfer learning model for online monitoring of surface roughness in milling with variable parameters

Surface roughness is crucial for the functional and aesthetic properties of mechanical components and must be carefully controlled during machining. However, predicting it under varying machining parameters is challenging due to limited experimental data and fluctuating factors like tool wear and vibration. This study develops a deep transfer learning model that incorporates the correlation alignment method and tool wear to enhance model generalization and reduce data acquisition costs. It utilizes multi-sensor data and the ResNet18 with a convolutional block attention module (CBAM-ResNet) to extract features with improved generalization and accuracy for monitoring milled surface roughness under varying conditions. The performance of the model is evaluated from different perspectives. First, the proposed model achieves high accuracy with fewer than 500 experimental samples from the target domain by using the CORAL module in the CBAM-ResNet model. This demonstrates the model's strong generalization capability by minimizing second-order statistical discrepancies between different datasets. Second, ablation experiments reveal a significant reduction in test error when incorporating CORAL and tool wear, highlighting their contributions to improved model generalization. Integrating tool wear information significantly reduces test errors across various transfer conditions, as it reflects changes in cutting force, vibration, and built-up edge formation. Third, comparisons with existing deep transfer models further emphasize the advantages of the proposed approach in improving model generalization. In summary, the proposed surface roughness model, which incorporates tool wear and multi-sensor signal features as inputs and employs feature transfer and CBAM-ResNet, demonstrates superior generalization and accuracy across various machining parameters.

Helmet Wearing Detection Algorithm Based on YOLOv5s-FCW

An enhanced algorithm, YOLOv5s-FCW, is put forward in this study to tackle the problems that exist in the current helmet detection (HD) methods. These issues include having too many parameters, a complex network, and large computation requirements, making it unsuitable for deployment on embedded and other devices. Additionally, existing algorithms struggle with detecting small targets and do not achieve high enough recognition accuracy. Firstly, the YOLOv5s backbone network is replaced by FasterNet for feature extraction (FE), which reduces the number of parameters and computational effort in the network. Secondly, a convolutional block attention module (CBAM) is added to the YOLOv5 model to improve the detection model’s ability to detect small objects such as helmets by increasing its attention to them. Finally, to enhance model convergence, the WIoU_Loss loss function is adopted instead of the GIoU_Loss loss function. As reported by the experimental results, the YOLOv5s-FCW algorithm proposed in this study has improved accuracy by 4.6% compared to the baseline algorithm. The proposed approach not only enhances detection concerning small and obscured targets but also reduces computation for the YOLOv5s model by 20%, thereby decreasing the hardware cost while maintaining a higher average accuracy regarding detection.

Road Defect Identification and Location Method Based on an Improved ML-YOLO Algorithm

The conventional method for detecting road defects relies heavily on manual inspections, which are often inefficient and struggle with precise defect localization. This paper introduces a novel approach for identifying and locating road defects based on an enhanced ML-YOLO algorithm. By refining the YOLOv8 object detection framework, we optimize both the traditional convolutional layers and the spatial pyramid pooling network. Additionally, we incorporate the Convolutional Block Attention to effectively capture channel and spatial features, along with the Selective Kernel Networks that dynamically adapt to feature extraction across varying scales. An optimized target localization algorithm is proposed to achieve high-precision identification and accurate positioning of road defects. Experimental results indicate that the detection accuracy of the improved ML-YOLO algorithm reaches 0.841, with a recall rate of 0.745 and an average precision of 0.817. Compared to the baseline YOLOv8 model, there is an increase in accuracy by 0.13, a rise in recall rate by 0.117, and an enhancement in average precision by 0.116. After the high detection accuracy of road defects was confirmed, generalization experiments were carried out on the improved ML-YOLO model in the public data set. The experimental results showed that compared with the original YOLOv8n, the average precision and recall rate of all types of ML-YOLO increased by 0.075, 0.121, and 0.035 respectively, indicating robust generalization capabilities. When applied to real-time road monitoring scenarios, this algorithm facilitates precise detection and localization of defects while significantly mitigating traffic accident risks and extending roadway service life. A high detection accuracy of road defects was achieved.

LT-DeepLab: an improved DeepLabV3+ cross-scale segmentation algorithm for Zanthoxylum bungeanum Maxim leaf-trunk diseases in real-world environments

IntroductionZanthoxylum bungeanum Maxim is an economically significant crop in Asia, but large-scale cultivation is often threatened by frequent diseases, leading to significant yield declines. Deep learning-based methods for crop disease recognition have emerged as a vital research area in agriculture.MethodsThis paper presents a novel model, LT-DeepLab, for the semantic segmentation of leaf spot (folium macula), rust, frost damage (gelu damnum), and diseased leaves and trunks in complex field environments. The proposed model enhances DeepLabV3+ with an innovative Fission Depth Separable with CRCC Atrous Spatial Pyramid Pooling module, which reduces the structural parameters of Atrous Spatial Pyramid Pooling module and improves cross-scale extraction capability. Incorporating Criss-Cross Attention with the Convolutional Block Attention Module provides a complementary boost to channel feature extraction. Additionally, deformable convolution enhances low-dimensional features, and a Fully Convolutional Network auxiliary header is integrated to optimize the network and enhance model accuracy without increasing parameter count.ResultsLT-DeepLab improves the mean Intersection over Union (mIoU) by 3.59%, the mean Pixel Accuracy (mPA) by 2.16%, and the Overall Accuracy (OA) by 0.94% compared to the baseline DeepLabV3+. It also reduces computational demands by 11.11% and decreases the parameter count by 16.82%.DiscussionThese results indicate that LT-DeepLab demonstrates excellent disease segmentation capabilities in complex field environments while maintaining high computational efficiency, offering a promising solution for improving crop disease management efficiency.

Spatial Resolution Enhancement Framework Using Convolutional Attention-Based Token Mixer

Spatial resolution enhancement in remote sensing data aims to augment the level of detail and accuracy in images captured by satellite sensors. We proposed a novel spatial resolution enhancement framework using the convolutional attention-based token mixer method. This approach leveraged spatial context and semantic information to improve the spatial resolution of images. This method used the multi-head convolutional attention block and sub-pixel convolution to extract spatial and spectral information and fused them using the same technique. The multi-head convolutional attention block can effectively utilize the local information of spatial and spectral dimensions. The method was tested on two kinds of data types, which were the visual-thermal dataset and the visual-hyperspectral dataset. Our method was also compared with the state-of-the-art methods, including traditional methods and deep learning methods. The experiment results showed that the method was effective and outperformed state-of-the-art methods in overall, spatial, and spectral accuracies.

Marine Oil Pollution Monitoring Based on a Morphological Attention U-Net Using SAR Images.

This study proposed an improved full-scale aggregated MobileUNet (FA-MobileUNet) model to achieve more complete detection results of oil spill areas using synthetic aperture radar (SAR) images. The convolutional block attention module (CBAM) in the FA-MobileUNet was modified based on morphological concepts. By introducing the morphological attention module (MAM), the improved FA-MobileUNet model can reduce the fragments and holes in the detection results, providing complete oil spill areas which were more suitable for describing the location and scope of oil pollution incidents. In addition, to overcome the inherent category imbalance of the dataset, label smoothing was applied in model training to reduce the model's overconfidence in majority class samples while improving the model's generalization ability. The detection performance of the improved FA-MobileUNet model reached an mIoU (mean intersection over union) of 84.55%, which was 17.15% higher than that of the original U-Net model. The effectiveness of the proposed model was then verified using the oil pollution incidents that significantly impacted Taiwan's marine environment. Experimental results showed that the extent of the detected oil spill was consistent with the oil pollution area recorded in the incident reports.

Prediction of lumpy skin disease virus using customized CBAM-DenseNet-attention model

Lumpy skin disease virus (LSDV) is an extremely infectious, viral, and chronic skin disease that is caused by the Capripox virus. This viral disease is predominantly found in cows. Mosquitoes and ticks are the primary transmitters for the spread of this virus. Recently, LSDV has been rapidly spreading all over the world, especially in several areas of Pakistan, India, and Iran. Thousands of cows have died due to this infectious virus in Pakistan and early detection of LSDV is needed to avoid further loss. The prediction and classification of LSDV are hindered by the lack of publicly available datasets. Despite a few studies using LSDV datasets, such datasets are often small, which may lead to model overfitting. In this regard, we collect the dataset from several online sources, as well as, collecting images from veterinary farms in different areas of Pakistan. Deep learning has been widely used in the medical field for disease detection and classification. Therefore, this study leverages DenseNet deep learning models for LSDV detection and classification. Experiments are performed using VGG-16, ResNet-50, MobileNet-V2, custom-designed convolutional neural network, and Inception-V3. The DenseNet architecture presents a Convolutional Block Attention Module (CBAM) and Spatial Attention (SA) for the prediction and classification of LSD. Results demonstrate that a 99.11% accuracy can be obtained on the augmented dataset while a 94.23% accuracy can be achieved with the original dataset for chicken pox, monkey pox, and LSDV. Comparison with state-of-the-art studies corroborates the superior performance of the proposed model.

Research on fault diagnosis of mechanical component in high-speed train based on one-dimensional convolutional block residual channel attention

The yaw damper and the rolling bearing are key components to ensure the safe and reliable operation of high-speed trains. However, fault identification in these components remains challenging, with traditional feature extraction methods often suffering from low diagnostic accuracy. To address these issues, a one-dimensional convolutional block residual channel attention (1DCBRCA) method for fault diagnosis of high-speed train components is introduced. Firstly, the convolutional layer is constructed for feature extraction, and the convolution block attention module is utilized for adaptive feature optimization in both channel and space dimensions. Subsequently, a residual neural network model is established, with channel attention added after each residual block to focus on critical feature information, which is used for adjusting the weight parameters. Furthermore, the proposed method was validated through damper fault experiments and bearing fault experiments. Comparative analysis with state-of-the-art methods demonstrated that the proposed 1DCBRCA approach achieves higher diagnostic accuracy, confirming its potential to enhance fault diagnosis in high-speed train components.

Dual convolution-transformer UNet (DCT-UNet) for organs at risk and clinical target volume segmentation in MRI for cervical cancer brachytherapy

Objective. MRI is the standard imaging modality for high-dose-rate brachytherapy of cervical cancer. Precise contouring of organs at risk (OARs) and high-risk clinical target volume (HR-CTV) from MRI is a crucial step for radiotherapy planning and treatment. However, conventional manual contouring has limitations in terms of accuracy as well as procedural time. To overcome these, we propose a deep learning approach to automatically segment OARs (bladder, rectum, and sigmoid colon) and HR-CTV from female pelvic MRI.Approach. In the proposed pipeline, a coarse multi-organ segmentation model first segments all structures, from which a region of interest is computed for each structure. Then, each organ is segmented using an organ-specific fine segmentation model separately trained for each organ. To account for variable sizes of HR-CTV, a size-adaptive multi-model approach was employed. For coarse and fine segmentations, we designed a dual convolution-transformer UNet (DCT-UNet) which uses dual-path encoder consisting of convolution and transformer blocks. To evaluate our model, OAR segmentations were compared to the clinical contours drawn by the attending radiation oncologist. For HR-CTV, four sets of contours (clinical + three additional sets) were obtained to produce a consensus ground truth as well as for inter/intra-observer variability analysis.Main results. DCT-UNet achieved dice similarity coefficient (mean ± SD) of 0.932 ± 0.032 (bladder), 0.786 ± 0.090 (rectum), 0.663 ± 0.180 (sigmoid colon), and 0.741 ± 0.076 (HR-CTV), outperforming other state-of-the-art models. Notably, the size-adaptive multi-model significantly improved HR-CTV segmentation compared to a single-model. Furthermore, significant inter/intra-observer variability was observed, and our model showed comparable performance to all observers. Computation time for the entire pipeline per subject was 12.59 ± 0.79 s, which is significantly shorter than the typical manual contouring time of >15 min.Significance. These experimental results demonstrate that our model has great utility in cervical cancer brachytherapy by enabling fast and accurate automatic segmentation, and has potential in improving consistency in contouring. DCT-UNet source code is available athttps://github.com/JHU-MICA/DCT-UNet.

Identification and Counting of Pirapitinga Piaractus brachypomus Fingerlings Fish Using Machine Learning.

Identifying and counting fish are crucial for managing stocking, harvesting, and marketing of farmed fish. Researchers have used convolutional networks for these tasks and explored various approaches to enhance network learning. Batch normalization is one technique that improves network stability and accuracy. This study aimed to evaluate machine learning for identifying and counting pirapitinga Piaractus brachypomus fry with different batch sizes. The researchers used one thousand photographic images of Pirapitinga fingerlings, labeled with bounding boxes. They trained the adapted convolutional network model with batch normalization layers added at the end of each convolution block. They set the training to one hundred and fifty epochs and tested batch sizes of 5, 10, and 20. Furthermore, they measured network performance using precision, recall, and mAP@0.5. Models with smaller batch sizes performed less effectively. The training with a batch size of 20 achieved the best performance, with a precision of 96.74%, recall of 95.48%, mAP@0.5 of 97.08%, and accuracy of 98%. This indicates that larger batch sizes improve accuracy in detecting and counting pirapitinga fry across different fish densities.

Enhancing bridge damage detection with Mamba-Enhanced HRNet for semantic segmentation.

With the acceleration of urbanization, bridges, as crucial infrastructure, their structural health and stability are paramount to public safety. This paper proposes Mamba-Enhanced HRNet for bridge damage detection. Mamba-Enhanced HRNet integrates the advantages of HRNet's multi-resolution parallel design and VMamba's visual state space model. By replacing the residual convolutional blocks in HRNet with a combination of VSS blocks and convolution, this model enhances the network's capability to capture global contextual information while maintaining computational efficiency. This work builds an extensive dataset with multiple damage kinds and uses Mean Intersection over Union (Mean IoU) as the assessment metric to assess the performance of Mamba-Enhanced HRNet. Experimental results demonstrate that Mamba-Enhanced HRNet achieves significant performance improvements in bridge damage semantic segmentation tasks, with Mean IoU scores of 0.963, outperforming several other semantic segmentation models.

A New Kiwi Fruit Detection Algorithm Based on an Improved Lightweight Network

To address the challenges associated with kiwi fruit detection methods, such as low average accuracy, inaccurate recognition of fruits, and long recognition time, this study proposes a novel kiwi fruit recognition method based on an improved lightweight network S-YOLOv4-tiny detection algorithm. Firstly, the YOLOv4-tiny algorithm utilizes the CSPdarknet53-tiny network as a backbone feature extraction network, replacing the CSPdarknet53 network in the YOLOv4 algorithm to enhance the speed of kiwi fruit recognition. Additionally, a squeeze-and-excitation network has been incorporated into the S-YOLOv4-tiny detection algorithm to improve accurate image extraction of kiwi fruit characteristics. Finally, enhancing dataset pictures using mosaic methods has improved precision in the characteristic recognition of kiwi fruits. The experimental results demonstrate that the recognition and positioning of kiwi fruits have yielded improved outcomes. The mean average precision (mAP) stands at 89.75%, with a detection precision of 93.96% and a single-picture detection time of 8.50 ms. Compared to the YOLOv4-tiny detection algorithm network, the network in this study exhibits a 7.07% increase in mean average precision and a 1.16% acceleration in detection time. Furthermore, an enhancement method based on the Squeeze-and-Excitation Network (SENet) is proposed, as opposed to the convolutional block attention module (CBAM) and efficient channel attention (ECA). This approach effectively addresses issues related to slow training speed and low recognition accuracy of kiwi fruit, offering valuable technical insights for efficient mechanical picking methods.

A dual-stream temporal convolutional network for remaining useful life prediction of rolling bearings

Abstract Remaining useful life (RUL) prediction plays an indispensable role in the reliable operation and improved maintenance of rolling bearings. Currently, data-driven methods based on deep learning have made significant progress in RUL prediction. However, most of such methods only consider the correlation between channels, ignoring the importance of different time steps for RUL prediction. In addition, it is still challenging to effectively fuse the degradation features of rolling bearings to improve the model RUL prediction performance. To address the above issues, this paper proposes a novel data-driven RUL prediction method named dual-stream temporal convolution network (DSTCN). First, a hybrid attention temporal convolution block (HATCB) is designed to capture the correlation of degraded features on the channel dimension and temporal dimension. Second, a one-dimensional attention fusion module is designed. This module is capable of weight recalibration and assignment to adaptively fuse different degraded features. Afterward, the Hilbert Marginal spectrum is obtained using the Hilbert Huang Transform and used as the input to one stream. Meanwhile, vibration signals are used as the input of the other stream, thus building a dual-stream temporal convolution network to realize RUL prediction. The effectiveness of the proposed method is validated with two life-cycle datasets, and the results show that the method has lower prediction error than other methods for RUL prediction and prognostic analysis.

Structural topology optimization based on deep learning

In the mechanical design of structures, traditional topology optimization methods involve numerous finite element iterative analyses, leading to a significant expenditure of computational resources. Therefore, the improved multi-scale gradient generative adversarial networks topology optimization technique is proposed. The topology optimization condition parameters are compressed into a low-dimensional latent space feature representation using the encoder, allowing the model to better extract features from these parameters. To speed up model training, the generator and discriminator networks use lightweight residual convolutional blocks. The hybrid attention mechanism extracts prominent region features from the topology optimization structure map. The model training process is guided by a multi-dimensional fusion loss function to enhance the quality of generated model samples. Finally, transferring the parameters of the low-resolution topology optimization model to the high-resolution model enables complete training on a limited amount of high-resolution topology optimization datasets. The experimental data on the low- and high-resolution topology optimization datasets demonstrate that, when compared to alternative methods, this method produces better-quality topology optimization structure maps. Additionally, it can generate high-resolution topology optimization structure maps in minimal time, enabling real-time topology optimization.

Visual tracking algorithm based on template updating and dual feature enhancement

Aiming at the problem of tracking failure due to target deformation, flipping and occlusion in visual tracking, a template updating algorithm based on image structural similarity is proposed by dynamically updating the template to adapt to the changes of the target during tracking, specifically, a queue is used to save the recent N -frame tracking results, and a decision is made on whether to update the template or not based on the structural similarity score between the current tracking results and the template image, and if updated, the template is matched from the historical N -frame tracking results as the new template. If updated, the optimal target image is matched from the historical N -frame tracking results as a new template for subsequent tracking. The tracking feature enhancement module and segmentation feature enhancement module are also designed based on SiamMask network. The tracking feature enhancement module consists of non-local operations and convolutional downsampling, which is used to establish contextual correlation, enhance the target features, suppress the background interference, improve the tracking robustness, and solve the feature attenuation problem due to the occlusion of the target. The segmentation feature enhancement module introduces the convolutional block attention module and deformable convolution to improve the network's ability to capture channel and spatial features, adaptively learn the shape and contour information of the target, and enhance the network's segmentation accuracy of the tracked target. , which in turn improves the tracking accuracy. Experiments show that the proposed algorithm performs well and stably in solving the above problems, improving the expected average overlap rate by 5.2%, 5.3%, and 2.5%, and the robustness by 6%, 7.9%, and 15.6%on the VOT2016, VOT2018, and VOT2019 datasets , respectively, and achieving a real-time speed of 91 frames per second, when compared to the baseline SiamMask.

Significant feature suppression and cross-feature fusion networks for fine-grained visual classification

The technique of extracting different distinguishing features by locating different part regions to achieve fine-grained visual classification (FGVC) has made significant improvements. Utilizing attention mechanisms for feature extraction has become one of the mainstream methods in computer vision, but these methods have certain limitations. They typically focus on the most discriminative regions and directly combine the features of these parts, neglecting other less prominent yet still discriminative regions. Additionally, these methods may not fully explore the intrinsic connections between higher-order and lower-order features to optimize model classification performance. By considering the potential relationships between different higher-order feature representations in the object image, we can enable the integrated higher-order features to contribute more significantly to the model’s classification decision-making capabilities. To this end, we propose a saliency feature suppression and cross-feature fusion network model (SFSCF-Net) to explore the interaction learning between different higher-order feature representations. These include (1) an object-level image generator (OIG): the intersection of the output feature maps of the last two convolutional blocks of the backbone network is used as an object mask and mapped to the original image for cropping to obtain an object-level image, which can effectively reduce the interference caused by complex backgrounds. (2) A saliency feature suppression module (SFSM): the most distinguishing part of the object image is obtained by a feature extractor, and the part is masked by a two-dimensional suppression method, which improves the accuracy of feature suppression. (3) A cross-feature fusion method (CFM) based on inter-layer interaction: the output feature maps of different network layers are interactively integrated to obtain high-dimensional features, and then the high-dimensional features are channel compressed to obtain the inter-layer interaction feature representation, which enriches the output feature semantic information. The proposed SFSCF-Net can be trained end-to-end and achieves state-of-the-art or competitive results on four FGVC benchmark datasets.