Large Receptive Fields Research Articles

SGFNet: Semantic-Guided Fusion Network for RGB-Thermal Semantic Segmentation

Recently, semantic segmentation based on RGB and thermal infrared (TIR) images has become a research hotspot because of its stability in the weak light environment. However, most of the current methods ignore the differences between the two modalities of data and do not use semantic information in multi-modal fusion. In this paper, we propose a novel Semantic-Guided Fusion Network (SGFNet) for RGB-Thermal semantic segmentation, which makes full use of semantic information in the multi-modal fusion. Our SGFNet consists of an asymmetric encoder with TIR branch and RGB branch and a decoder. We concentrate on enhancing the multi-modal feature representation in the encoder with a pattern of fusion and enhancement. Specifically, considering that TIR images are stable under weak light conditions, we first propose a Semantic Guidance Head to extract semantic information in the TIR branch. In the RGB branch, we propose a Multi-modal Coordination and Distillation Unit to fuse multi-modal features first. Then, we propose a Cross-level and Semantic-guided Enhancement Unit to enhance the fused features with cross-level information and semantic information. We arrange these two units at all stages of the RGB branch to generate features with strong representation abilities at different levels. For the decoder, to obtain large receptive fields and fine edges, we improve the Lawin ASPP decoder by introducing edge information extracted from the low-level features, proposing the edge-aware Lawin ASPP decoder. With our encoder and decoder working together, our SGFNet can identify objects accurately and segment objects finely. Extensive experiments on the MFNet dataset demonstrate the superior performance of the proposed SGFNet compared with state-of-the-art methods. The code and results of our method are available at https://github.com/kw717/SGFNet.

Quantifying spatial acuity of frequency resolved midair ultrasound vibrotactile stimuli

Spatial acuity is a fundamental property of any sensory system. In the case of the somatosensory system, the two-point discrimination (2PD) test has long been used to investigate tactile spatial resolution. However, the somatosensory system comprises three main mechanoreceptive channels: the slowly adapting channel (SA) responds to steady pressure, the rapidly adapting channel (RA) responds to low-frequency vibration, and the Pacinian channel (PC) responds to high-frequency vibration. The use of mechanical stimuli in the classical 2PD test means that previous studies on tactile acuity have primarily focussed on the pressure-sensitive channel alone, while neglecting other submodalities. Here, we used a novel ultrasound stimulation to systematically investigate the spatial resolution of the two main vibrotactile channels. Contrary to the textbook view of poor spatial resolution for PC-like stimuli, across four experiments we found that high-frequency vibration produced surprisingly good spatial acuity. This effect remained after controlling for interchannel differences in stimulus detectability and perceived intensity. Laser doppler vibrometry experiments confirmed that the acuity of the PC channel was not simply an artifact of the skin’s resonance to high-frequency mechanical stimulation. Thus, PC receptors may transmit substantial spatial information, despite their sparse distribution, deep location, and large receptive fields.

Quadtree decomposition-based Deep learning method for multiscale coastline extraction with high-resolution remote sensing imagery

As one of the most critical features on the earth's surface, coastal zone mandates high-quality extraction of its representative feature, the coastline. Prior methodologies primarily emphasize on edge and small-scale information. However, during large-scale image processing, misclassification might occur due to the difficulty in determining whether a local area belongs to the land or sea. To address this, we propose a deep learning-based multiscale coastline extraction algorithm in this study. It comprises a multiscale coastal zone dataset built upon a tile map service structure and a scene classification-based multiscale coastal zone classifier, employing quadtree decomposition to identify coastal zones from low to high levels. Contrasting with conventional semantic segmentation, the scene classification network, owing to its larger receptive field, can accurately discern land and sea. This accuracy is further enhanced by using quadtree decomposition to process images with lower resolution and larger coverage. The results suggest that our proposed method effectively eliminates confusing features, with the overall experimental classification accuracy attesting to the effectiveness of our approach, yielding a 6% improvement. Moreover, the screening process in this study significantly reduces the number of input samples for the segmentation network, thus boosting computational speed.

Real-Time Dynamic Gesture Recognition Algorithm Based on Adaptive Information Fusion and Multi-Scale Optimization Transformer

Gesture recognition is a popular technology in the field of computer vision and an important technical mean of achieving human-computer interaction. To address problems such as the limited long-range feature extraction capability of existing dynamic gesture recognition networks based on convolutional operators, we propose a dynamic gesture recognition algorithm based on spatial pyramid pooling Transformer and optical flow information fusion. We take advantage of Transformer’s large receptive field to reduce model computation while improving the model’s ability to extract features at different scales by embedding spatial pyramid pooling. We use the optical flow algorithm with the global motion aggregation module to obtain an optical flow map of hand motion, and to extract the key frames based on the similarity minimization principle. We also design an adaptive feature fusion method to fuse the spatial and temporal features of the dual channels. Finally, we demonstrate the effectiveness of model components on model recognition enhancement through ablation experiments. We conduct training and validation on the SCUT-DHGA dynamic gesture dataset and on a dataset we collected, and we perform real-time dynamic gesture recognition tests using the trained model. The results show that our algorithm achieves high accuracy even while keeping the parameters balanced. It also achieves fast and accurate recognition of dynamic gestures in real-time tests.

A lightweight efficient semantic segmentation with encoder-decoder for arc interference in robotic arc welding

Recognizing laser stripes under conditions of arc interference is a persistent challenge faced by linear structured light sensors due to their similarity with noise in morphological characteristics. In this paper, a symmetrical encoder-decoder semantic segmentation model named LSRNet (lightweight stripe recognition network) is proposed that enables real-time detection of laser stripes in environments with strong arc interference. It analyzes the difference between arc and laser stripe from the semantic point of view. Firstly, to facilitate quick and accurate recognition of laser stripes, we designed a Depthwise-separable Factorized Module that is both efficient and scalable with a larger receptive field. Moreover, due to the difficulty in distinguishing laser stripes from arcs in a local context, a Multi-scale Depthwise-separable Factorized Module is designed to capture more context information. Secondly, the Unified Attention Fusion Module is used to enhance the feature representation of the laser stripe. Additionally, The combination of DiceLoss(dice coefficient loss) and CELoss(cross-entropy loss) is introduced to optimize the loss, which addresses the pixel imbalance challenge that the pixels of the laser stripe are much less than the background pixels. Finally, based on the images collected in the industrial field, our ArcDataset is created for model training. Experimental results show that the proposed method achieves the highest intersection over union of 88.0% in laser stripe segmentation and a recognition speed of 101.3 FPS(frames per second).

Data-driven algorithm based on the scaled boundary finite element method and deep learning for the identification of multiple cracks in massive structures

Structural defect identification is a vital aspect of structural health monitoring used to assess the safety of engineering structures. However, quantitatively determining the dimensions of structural defects is often difficult. Therefore, this study presents an innovative data-driven algorithm that combines the scaled boundary finite element method (SBFEM) and a deep learning framework based on a dilated causal convolutional neural network (CNN) to identify crack-like defects in large-scale structures. The SBFEM is used to simulate different crack-like defects. Mesh generation is significantly simplified by a simple procedure that requires only changing the scale centre at the crack tip and the positions of the nodes at the crack opening. This minimises remeshing and enables simple generation of sufficient data to train the neural network. In addition, an absorbing boundary model based on Rayleigh damping is used to avoid computing the entire model when simulating wave propagation in massive structures. To ensure that sequential data remain ordered and to obtain a large receptive field without increasing the complexity of the neural network, a dilated causal CNN is employed in the deep learning framework. Therefore, more historical information is captured, and the complex mapping relationship between the echo signal and the crack information is efficiently learnt. The proposed model can accurately identify the number, location, and depth of cracks in massive structures. Moreover, it is robust to noise, which is demonstrated via numerical examples. Therefore, the proposed algorithm provides valuable insight into the detection and diagnosis of structural defects, which can ultimately improve the safety of engineering structures.

A dilated convolution-based feature adaptation method for detection of high aspect ratio objects in aerial images

In real scenarios, objects with high aspect ratios are actually very common, and such objects hold significant importance in the field of object detection. However, most of the existing object detection algorithms tend to overlook this specific type of object. After analyzing the statistical data, we observed a substantial decrease in mAP (mean Average Precision) for classical object detection algorithms when they are tasked with detecting only high aspect ratio objects. Therefore, we conducted an analysis of the factors that influence the detection performance of these objects and made the following improvements: (1) We introduced large-kernel attention convolution between the backbone network layers. This addition allows each position feature to have a larger receptive field, facilitating better feature learning; (2) By incorporating multiple sets of deformable convolutions for feature-adaptive processing, we were able to enhance the learning of characteristic information specific to the object itself. This approach also promotes network convergence. The proposed method yielded a significant improvement in accuracy, approximately 5[Formula: see text] higher than the baseline, when evaluated on the FGSD2021 dataset. Furthermore, our method outperformed the current best method by approximately 0.5[Formula: see text].

A dual-path instance segmentation network based on nuclei contour in histology image

Accurate segmentation of nuclei in histology images is essential for digital pathology. However, previous work should have considered integrating nuclei contour information into network learning. The main problems are (1) nuclei contour information needs to be fully considered, resulting in inaccurate spatial location of nuclei. (2) Small nuclei in dense regions or irregularly shaped nuclei cannot be segmented. (3) Mainstream networks consider more long-distance semantic information and larger receptive fields and do not consider the fusion of feature maps with different semantics. To solve the above problems, we have proposed a contour-based dual-path instance segmentation network. Firstly we reconstructed the nuclei contour information using algorithms and morphological manipulations. Provide input for subsequent contouring networks. Then we designed a dual-path network. It can extract nuclei features and contour features independently in the encoding stage and fuse the feature maps at different scales in the decoding stage. In the decoding stage, we use the attention module with a newly designed fusion mechanism, which allows us to fuse different semantics of feature maps for simple and efficient fusion. Finally, using the watershed algorithm, we use the network segmentation results to get the instance segmentation results of the nuclei. We used four publicly available datasets, including Kumar, CPM-15, CPM-17, and TNBC, which contain the nuclei of many organs and different forms of nuclei in the human body. Compared with the mainstream methods, we obtained the best AJI metrics, which are 0.656,0.717, and 0.602. In conclocess can effectively improve the performance of the nuclei instance segmentation network. It can accurately locate the scattered nuclei in space with small segmented nuclei in dense regions.

Xoo-YOLO: a detection method for wild rice bacterial blight in the field from the perspective of unmanned aerial vehicles.

Wild rice, a natural gene pool for rice germplasm innovation and variety improvement, holds immense value in rice breeding due to its disease-resistance genes. Traditional disease resistance identification in wild rice heavily relies on labor-intensive and subjective manual methods, posing significant challenges for large-scale identification. The fusion of unmanned aerial vehicles (UAVs) and deep learning is emerging as a novel trend in intelligent disease resistance identification. Detecting diseases in field conditions is critical in intelligent disease resistance identification. In pursuit of detecting bacterial blight in wild rice within natural field conditions, this study presents the Xoo-YOLO model, a modification of the YOLOv8 model tailored for this purpose. The Xoo-YOLO model incorporates the Large Selective Kernel Network (LSKNet) into its backbone network, allowing for more effective disease detection from the perspective of UAVs. This is achieved by dynamically adjusting its large spatial receptive field. Concurrently, the neck network receives enhancements by integrating the GSConv hybrid convolution module. This addition serves to reduce both the amount of calculation and parameters. To tackle the issue of disease appearing elongated and rotated when viewed from a UAV perspective, we incorporated a rotational angle (theta dimension) into the head layer's output. This enhancement enables precise detection of bacterial blight in any direction in wild rice. The experimental results highlight the effectiveness of our proposed Xoo-YOLO model, boasting a remarkable mean average precision (mAP) of 94.95%. This outperforms other models, underscoring its superiority. Our model strikes a harmonious balance between accuracy and speed in disease detection. It is a technical cornerstone, facilitating the intelligent identification of disease resistance in wild rice on a large scale.

Convolutional neural networks rarely learn shape for semantic segmentation

Shape learning, or the ability to leverage shape information, could be a desirable property of convolutional neural networks (CNNs) when target objects have specific shapes. While some research on the topic is emerging, there is no systematic study to conclusively determine whether and under what circumstances CNNs learn shape. Here, we present such a study in the context of segmentation networks where shapes are particularly important. We define shape and propose a new behavioral metric to measure the extent to which a CNN utilizes shape information. We then execute a set of experiments with synthetic and real-world data to progressively uncover under which circumstances CNNs learn shape and what can be done to encourage such behavior. We conclude that (i) CNNs do not learn shape in typical settings but rather rely on other features available to identify the objects of interest, (ii) CNNs can learn shape, but only if the shape is the only feature available to identify the object, (iii) sufficiently large receptive field size relative to the size of target objects is necessary for shape learning; (iv) a limited set of augmentations can encourage shape learning; (v) learning shape is indeed useful in the presence of out-of-distribution data.

Atrial fibrillation detection with signal decomposition and dilated residual neural network

Objective. Detecting atrial fibrillation (AF) using electrocardiogram (ECG) recordings from wearable devices has been challenging due to factors such as low signal-to-noise ratio and the use of only one lead. The use of deep learning has become a popular approach to tackle this task. However, it has been observed that current methods based on deep neural networks tend to favor raw signals as input, disregarding the valuable clinical experience in ECG diagnosis. Approach. In this study, we proposed a novel feature extraction method that generates a pseudo QRS complex signal and a pseudo T, P wave signal for each raw ECG signal using a temporal mask built upon R peak detection. Then a novel dilated residual neural network was trained on the decomposed signal. Main results. We evaluated the performance of our method on the dataset of PhysioNet/CinC 2017 Challenge, achieving an average score of 0.843. The method was further tested on MIT-BIH Atrial Fibrillation Database, and an average score of 0.984 was obtained. Significance. Our proposed ECG signal decomposition technique introduces simple and reliable domain knowledge into deep neural networks, and the dilated residual network provides large and flexible receptive fields, thereby enhancing the performance in the detection of AF. Our method can be extended to many other tasks involving ECG signals.

Transformer With Double Enhancement for Low-Dose CT Denoising.

Increasingly serious health problems have made the usage of computed tomography surge. Therefore, algorithms for processing CT images are becoming more and more abundant. These algorithms can lessen the harm of cumulative radiation in CT technology for the patient while eliminating the noise of image caused by dose reduction. However, the mainstream CNN-based algorithms are inefficient when dealing with features in broad regions. Inspired by the large receptive field of transformer framework, this paper designs an end-to-end low-dose CT (LDCT) denoising network based on the transformer. The overall framework of network contains a main branch and dual side branches. Specifically, the overlapping-free window-based self-attention transformer block is adopted on the main branch to realize image denoising. On the dual side branches, we propose double enhancement module to enrich edge, texture, and context information of LDCT images. Meanwhile, the receptive field of network is further enlarged after processing, which is helpful for building model's long-range dependencies. The outputs of the side branches are concatenated to the middle layer of the network for enhancing information and generating high-quality CT images. In addition, to better train the network and achieve the expected results, we introduce a compound loss function including mean squared error (MSE), multi-scale perceptual (MSP), and Sobel-L1 (SL) to make the denoised image closer to the targeted norm-dose CT (NDCT) image. Lastly, we conducted experiments on two clinical datasets including abdomen, head, and chest LDCT images with 25%, 25%, and 10% of the full dose, respectively. The experimental results demonstrated that the proposed DEformer achieved better denoising performance than the existing algorithms.

Poster Session: Towards Understanding Retinal Processing of Single-Cone Scale Stimulation.

The magnocellular (M) visual pathway can be distinguished from its finer-grained parvocellular (P) counterpart by its stronger response gain to contrasts near human detection threshold. This distinction is found when stimulus size is matched to the relatively large receptive field size of M cells, but it is less clear if preserved at smaller spatial scales. We examined foveal detection thresholds in tandem with a masking paradigm known to desensitize the M pathway. We used a custom DLP display in Maxwellian view to present either 3 or 23 arcmin spots for 100 ms against a uniform gray disk (1° VA) surrounded by a 0.5 cpd, 100% contrast, square wave grating flickered at 7 frequencies ranging from 0 to 30 Hz. Detection data were fit with psychometric functions to estimate threshold and thresholds were normalized to the static (0 Hz) condition. Relative sensitivity was fit with a modified impulse response function and compared to a constant model with ANOVA. With masking we found a significant reduction in sensitivity for large (23arcmin) increments (mean=0.39 log units; n=4; p<0.01) and decrements (0.31 log units; n=4; p=0.02)at 8.3Hz. By contrast, no significant masking effect was observed for small spot increments or decrements (p>0.05). These data suggest that the M pathway is not the primary determinant of small spot thresholds under the conditions studied. Further investigation with adaptive optics may be necessary to fully elaborate single-cone processing in the fovea.

HyFormer: Hybrid Grouping-Aggregation Transformer and Wide-Spanning CNN for Hyperspectral Image Super-Resolution

Hyperspectral image (HSI) super-resolution is a practical and challenging task as it requires the reconstruction of a large number of spectral bands. Achieving excellent reconstruction results can greatly benefit subsequent downstream tasks. The current mainstream hyperspectral super-resolution methods mainly utilize 3D convolutional neural networks (3D CNN) for design. However, the commonly used small kernel size in 3D CNN limits the model’s receptive field, preventing it from considering a wider range of contextual information. Though the receptive field could be expanded by enlarging the kernel size, it results in a dramatic increase in model parameters. Furthermore, the popular vision transformers designed for natural images are not suitable for processing HSI. This is because HSI exhibits sparsity in the spatial domain, which can lead to significant computational resource waste when using self-attention. In this paper, we design a hybrid architecture called HyFormer, which combines the strengths of CNN and transformer for hyperspectral super-resolution. The transformer branch enables intra-spectra interaction to capture fine-grained contextual details at each specific wavelength. Meanwhile, the CNN branch facilitates efficient inter-spectra feature extraction among different wavelengths while maintaining a large receptive field. Specifically, in the transformer branch, we propose a novel Grouping-Aggregation transformer (GAT), comprising grouping self-attention (GSA) and aggregation self-attention (ASA). The GSA is employed to extract diverse fine-grained features of targets, while the ASA facilitates interaction among heterogeneous textures allocated to different channels. In the CNN branch, we propose a Wide-Spanning Separable 3D Attention (WSSA) to enlarge the receptive field while keeping a low parameter number. Building upon WSSA, we construct a wide-spanning CNN module to efficiently extract inter-spectra features. Extensive experiments demonstrate the superior performance of our HyFormer.

Multiple Axes of Visual Symmetry: Detection and Aesthetic Preference

Little is known about the detection of and preference for multiple simultaneous parallel axes of symmetry. Neuroscientists have suggested that the detection of symmetry occurs in extrastriate brain areas with large receptive fields. Such large receptive fields may potentially hinder the simultaneous detection of more than one axis of symmetry. In contrast, psychophysicists have found that symmetry detection occurs within small spatial windows, allowing for the concurrent detection of multiple axes of symmetry. Using psychophysical and computational methods, we aim to test whether multiple axes of symmetry can be detected in parallel and to understand the role of multiple axes of symmetry on aesthetic valence. Experiment 1 provides evidence that multiple axes of symmetry cannot be detected simultaneously. However, with relatively long temporal integration, people can detect them. Experiment 2 suggests that multiple axes of symmetry tend to increase preference. However, the preference for symmetry is not universal because, although most people prefer symmetry, others prefer complex images without axes of symmetry. We present and test a computational model that explains the results of these experiments.

Drone-YOLO: An Efficient Neural Network Method for Target Detection in Drone Images

Object detection in unmanned aerial vehicle (UAV) imagery is a meaningful foundation in various research domains. However, UAV imagery poses unique challenges, including large image sizes, small sizes detection objects, dense distribution, overlapping instances, and insufficient lighting impacting the effectiveness of object detection. In this article, we propose Drone-YOLO, a series of multi-scale UAV image object detection algorithms based on the YOLOv8 model, designed to overcome the specific challenges associated with UAV image object detection. To address the issues of large scene sizes and small detection objects, we introduce improvements to the neck component of the YOLOv8 model. Specifically, we employ a three-layer PAFPN structure and incorporate a detection head tailored for small-sized objects using large-scale feature maps, significantly enhancing the algorithm’s capability to detect small-sized targets. Furthermore, we integrate the sandwich-fusion module into each layer of the neck’s up–down branch. This fusion mechanism combines network features with low-level features, providing rich spatial information about the objects at different layer detection heads. We achieve this fusion using depthwise separable evolution, which balances parameter costs and a large receptive field. In the network backbone, we employ RepVGG modules as downsampling layers, enhancing the network’s ability to learn multi-scale features and outperforming traditional convolutional layers. The proposed Drone-YOLO methods have been evaluated in ablation experiments and compared with other state-of-the-art approaches on the VisDrone2019 dataset. The results demonstrate that our Drone-YOLO (large) outperforms other baseline methods in the accuracy of object detection. Compared to YOLOv8, our method achieves a significant improvement in mAP0.5 metrics, with a 13.4% increase on the VisDrone2019-test and a 17.40% increase on the VisDrone2019-val. Additionally, the parameter-efficient Drone-YOLO (tiny) with only 5.25 M parameters performs equivalently or better than the baseline method with 9.66M parameters on the dataset. These experiments validate the effectiveness of the Drone-YOLO methods in the task of object detection in drone imagery.

Hybrid CNN-Transformer model for medical image segmentation with pyramid convolution and multi-layer perceptron

Vision Transformer (ViT) has emerged as a potential alternative to convolutional neural networks for large datasets. However, applying ViT directly to medical image segmentation is challenging due to its lack of induction bias, which requires a large number of high-quality annotated medical images for effective model training. Recent studies have discovered that, in addition to the increased model capacity and generalization resulting from the lack of induction bias, the excellent performance of Transformer can also be attributed to its large receptive field. In this paper, we propose a U-shaped medical image segmentation model that combines large kernel convolutions with Transformers. Specifically, we construct a basic Transformer unit using pyramidal convolution modules with multi-scale kernels and multi-layer perceptron. In the pyramid convolution module, we employ grouped convolution to reduce parameter and computational complexity while utilizing multi-scale large kernel attention as a foundation for more efficient feature extraction. For different types of grouping, different sizes of convolutions are used to enhance the extraction of features with multiple receptive fields. To optimize the extracted features from the encoder, the U-shaped model integrates a variant of the pyramidal convolutional module into the skip connections. This variant utilizes multi-scale large kernel convolutional attention based on channel splitting. The incorporation of this variant enables efficient refinement of the feature representations within the skip connections. Through extensive comparisons on multi-modal medical image datasets, our model outperforms state-of-the-art methods across various evaluation metrics, with notable superiority observed on small-scale medical datasets. Our research findings suggest that the combination of large kernel convolutions and Transformer models introduces an advantageous inductive bias, resulting in enhanced performance specifically for small-scale medical image datasets. To facilitate accessibility, we have made our code openly accessible on our GitHub repository, which can be found at https://github.com/medical-images-process/CNN-Transformer.

Effect of image resolution on automated classification of chest X-rays.

Deep learning (DL) models have received much attention lately for their ability to achieve expert-level performance on the accurate automated analysis of chest X-rays (CXRs). Recently available public CXR datasets include high resolution images, but state-of-the-art models are trained on reduced size images due to limitations on graphics processing unit memory and training time. As computing hardware continues to advance, it has become feasible to train deep convolutional neural networks on high-resolution images without sacrificing detail by downscaling. This study examines the effect of increased resolution on CXR classification performance. We used the publicly available MIMIC-CXR-JPG dataset, comprising 377,110 high resolution CXR images for this study. We applied image downscaling from native resolution to , , , and and then we used the DenseNet121 and EfficientNet-B4 DL models to evaluate clinical task performance using these four downscaled image resolutions. We find that while some clinical findings are more reliably labeled using high resolutions, many other findings are actually labeled better using downscaled inputs. We qualitatively verify that tasks requiring a large receptive field are better suited to downscaled low resolution input images, by inspecting effective receptive fields and class activation maps of trained models. Finally, we show that stacking an ensemble across resolutions outperforms each individual learner at all input resolutions while providing interpretable scale weights, indicating that diverse information is extracted across resolutions. This study suggests that instead of focusing solely on the finest image resolutions, multi-scale features should be emphasized for information extraction from high-resolution CXRs.

UcUNet: A lightweight and precise medical image segmentation network based on efficient large kernel U-shaped convolutional module design

In recent years, precise medical image segmentation methods based on the encoder–decoder structure have attracted much attention, but there are still some limitations. They mainly manifest in two aspects: one is that the network structure becomes increasingly complex in pursuit of segmentation accuracy, and the other is the insufficient multiscale information fusion ability between the encoder and decoder. To address these issues, we designed a novel lightweight precise medical image segmentation network called UcUNet, which has not only a very large receptive field and multiscale information fusion ability but also a low parameter count. Specifically, we first designed an efficient U-shaped convolution module that can increase the network depth with fewer parameters and effectively filter out invalid features while fusing shallow and deep features. Furthermore, we ingeniously introduced large-kernel convolution into this module, which significantly improved the network’s receptive field with very few parameters, providing new ideas for the use of large-kernel convolution in lightweight models. In addition, we designed a self-adjusting multiscale feature fusion module based on large-kernel convolution to replace the original simple skip connections, effectively enhancing the network’s multiscale information extraction and fusion ability. Finally, we conducted experiments on multiple datasets, and the results showed that our method can achieve precise medical image segmentation with a low parameter count.

SSD-Resnet50: Research on Pedestrian Detection Technology

In the era of artificial intelligence, with the rapid development of computer hardware, the defects of neural network algorithm have been further improved, the technology of artificial intelligence is progressing, the algorithm and technology of target detection have been further improved, and pedestrian detection technology is an important research direction of target detection technology. The main research content of this paper consists of three parts: the improvement of the original SSD algorithm, and the fundamental composition and characteristics of the neural network; Changing VGG16, the basis of the original SSD algorithm, to Resnet50 can improve the detection speed and accuracy. A feature fusion module is introduced to fuse the feature map with large receptive field and small receptive field, so as to improve the overall expression ability of the feature map to the image.