Effective Receptive Field Research Articles

Demystifying the effect of receptive field size in U-Net models for medical image segmentation.

Medical image segmentation is a critical task in healthcare applications, and U-Nets have demonstrated promising results in this domain. We delve into the understudied aspect of receptive field (RF) size and its impact on the U-Net and attention U-Net architectures used for medical imaging segmentation. We explore several critical elements including the relationship among RF size, characteristics of the region of interest, and model performance, as well as the balance between RF size and computational costs for U-Net and attention U-Net methods for different datasets. We also propose a mathematical notation for representing the theoretical receptive field (TRF) of a given layer in a network and propose two new metrics, namely, the effective receptive field (ERF) rate and the object rate, to quantify the fraction of significantly contributing pixels within the ERF against the TRF area and assessing the relative size of the segmentation object compared with the TRF size, respectively. The results demonstrate that there exists an optimal TRF size that successfully strikes a balance between capturing a wider global context and maintaining computational efficiency, thereby optimizing model performance. Interestingly, a distinct correlation is observed between the data complexity and the required TRF size; segmentation based solely on contrast achieved peak performance even with smaller TRF sizes, whereas more complex segmentation tasks necessitated larger TRFs. Attention U-Net models consistently outperformed their U-Net counterparts, highlighting the value of attention mechanisms regardless of TRF size. These insights present an invaluable resource for developing more efficient U-Net-based architectures for medical imaging and pave the way for future exploration of other segmentation architectures. A tool is also developed, which calculates the TRF for a U-Net (and attention U-Net) model and also suggests an appropriate TRF size for a given model and dataset.

Numerical investigation of the effective receptive field and its relationship with convolutional kernels and layers in convolutional neural network

The receptive field (RF) plays a crucial role in convolutional neural networks (CNNs) because it determines the amount of input information that each neuron in a CNN can perceive, which directly affects the feature extraction ability. As the number of convolutional layers in CNNs increases, there is a decay of the RF according to the two-dimensional Gaussian distribution. Thus, an effective receptive field (ERF) can be used to characterize the available part of the RF. The ERF is calculated by the kernel size and layer number within the neural network architecture. Currently, ERF calculation methods are typically applied to single-channel input data that are both independent and identically distributed. However, such methods may result in a loss of effective information if they are applied to more general (i.e., multi-channel) datasets. Therefore, we proposed a multi-channel ERF calculation method. By conducting a series of numerical experiments, we determined the relationship between the ERF and the convolutional kernel size in conjunction with the layer number. To validate the new method, we used the recently published global wave surrogate model for climate simulation (GWSM4C) and its accompanying dataset. According to the newly established relationship, we refined the kernel size and layer number in each neural network of the GWSM4C to produce the same ERF but lower RF attenuation rates than those of the original version. By visualizing the gradient map at several points in West African and East Pacific areas, the high gradient value regions confirmed the known swell sources, which indicated effective feature extraction in these areas. Furthermore, the new version of the GWSM4C yielded better prediction accuracy for significant wave height in global swell pools. The root mean square errors in the West African and East Pacific regions reduced from approximately 0.3 m, in the original model to about 0.15 m, in the new model. Moreover, these improvements were attributed to the higher efficiency of the newly modified neural network structure that allows the inclusion of more historical winds while maintaining acceptable computational consumption.

A generalized model for accurate wheat spike detection and counting in complex scenarios

Wheat is a crucial crop worldwide, and accurate detection and counting of wheat spikes are vital for yield estimation and breeding. However, these tasks are daunting in complex field environments. To tackle this, we introduce RIA-SpikeNet, a model designed to detect and count wheat spikes in such conditions. First, we introduce an Implicit Decoupling Detection Head to incorporate more implicit knowledge, enabling the model to better distinguish visually similar wheat spikes. Second, Asymmetric Loss is employed as the confidence loss function, enhancing the learning weights of positive and hard samples, thus improving performance in complex scenes. Lastly, the backbone network is modified through reparameterization and the use of larger convolutional kernels, expanding the effective receptive field and improving shape information extraction. These enhancements significantly improve the model’s ability to detect and count wheat spikes accurately. RIA-SpikeNet outperforms the state-of-the-art YOLOv8 detection model, achieving a competitive 81.54% mAP and 90.29% R2. The model demonstrates superior performance in challenging scenarios, providing an effective tool for wheat spike yield estimation in field environments and valuable support for wheat production and breeding efforts.

Multi-scale and multi-receptive field-based feature fusion for robust segmentation of plant disease and fruit using agricultural images

The accurate and fast assessment of plant diseases and fruits is important for sustainable and productive agriculture. However, typically manual methods are adopted for the assessment of plant diseases and fruit, which is a time-consuming, resource-intensive, and error-prone approach. A few artificial intelligence (AI)-based methods have been introduced to automate this process, but they show limitations in attaining high performance in challenging imaging conditions such as occlusions, poor lighting, noise, indistinctive boundaries, and excessive morphological variations. Moreover, current approaches also lack in providing computationally efficient solutions. To overcome these issues, two novel architectures are developed to segment plant diseases and fruits with higher segmentation performance and less computational requirements. The efficient feature fusion segmentation network (EFFS-Net) is the base network, and a multi-scale dilated feature fusion segmentation network (MDFS-Net) is the final network of this study. EFFS-Net uses an identity skip path-based feature fusion mechanism with an efficient grouped convolutional depth (EGCD) to provide satisfactory segmentation performance with high computational efficiency. MDFS-Net uses a multi-scale feature fusion mechanism and fuses low-level information in the EGCD and other sections of the network to learn detailed input features for delivering promising performance even in challenging imaging conditions. MDFS-Net also applies multiple receptive fields to the low-level information in the effective receptive field processing (ERFP) block for fusion near the pixel classification stage for further performance enhancement.Both networks are evaluated using a Brazilian Arabica coffee leaf (BRACOL) image dataset, an Australian Center for Field Robotics orchard fruit (apple) dataset, and a necrotized cassava root cross-section image dataset. The proposed method provides a promising segmentation performance, achieving dice similarity coefficients of 88.81 %, 95.01 %, and 86.04 % with performance improvement of 3.5 %, 2.6 %, and 1.07 % on the three datasets, respectively, with approximately ten times less number of required trainable parameters compared with the state-of-the-art methods.

Defect detection of printed circuit board assembly based on YOLOv5

Detection of printed circuit board assembly (PCBA) defects is crucial for improving the efficiency of PCBA manufacturing. This paper proposes PCBA-YOLO, a YOLOv5-based method that can effectively increase the accuracy of PCBA defect detection. First, the spatial pyramid pooling module with cross-stage partial structure is replaced in the neck network of YOLOv5 to capture the resolution features at multiple scales. Second, large kernel convolution is introduced in the backbone network to obtain larger effective receptive fields while reducing computational overhead. Finally, an SIoU loss function that considers the angular cost is utilized to enhance the convergence speed of the model. In addition, an assembled PCBA defect detection dataset named PCBA-DET is created in this paper, containing the corresponding defect categories and annotations of defect locations. The experimental results on the PCB defect dataset demonstrate that the improved method has lower loss values and higher accuracy. Evaluated on the PCBA-DET dataset, the mean average precision reaches 97.3%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\%$$\\end{document}, achieving a real-time detection performance of 322.6 frames per second, which considers both the detection accuracy and the model size compared to the YOLO series of detection networks. The source code and PCBA-DET dataset can be accessed at https://github.com/ismh16/PCBA-Dataset.

Multi-scale large kernel convolution and hybrid attention network for remote sensing image dehazing

Remote sensing (RS) image dehazing holds significant importance in enhancing the quality and information extraction capability of RS imagery. The enhancement in image dehazing quality has progressively advanced alongside the evolution of convolutional neural network (CNN). Due to the fixed receptive field of CNN, there is insufficient utilization of contextual information on haze features in multi-scale RS images. Additionally, the network fails to adequately extract both local and global information of haze features. In addressing the above problems, in this paper, we propose an RS image dehazing network based on multi-scale large kernel convolution and hybrid attention (MKHANet). The network is mainly composed of multi-scale large kernel convolution (MSLKC) module, hybrid attention (HA) module and feature fusion attention (FFA) module. The MSLKC module fully fuses the multi-scale information of features while enhancing the effective receptive field of the network by parallel multiple large kernel convolutions. To alleviate the problem of uneven distribution of haze and effectively extract the global and local information of haze features, the HA module is introduced by focusing on the importance of haze pixels at the channel level. The FFA module aims to boost the interaction of feature information between the network's deep and shallow layers. The subjective and objective experimental results on on multiple RS hazy image datasets illustrates that MKHANet surpasses existing state-of-the-art (SOTA) approaches. The source code is available at https://github.com/tohang98/MKHA_Net.

MixImages: An Urban Perception AI Method Based on Polarization Multimodalities.

Intelligent urban perception is one of the hot topics. Most previous urban perception models based on semantic segmentation mainly used RGB images as unimodal inputs. However, in natural urban scenes, the interplay of light and shadow often leads to confused RGB features, which diminish the model's perception ability. Multimodal polarization data encompass information dimensions beyond RGB, which can enhance the representation of shadow regions, serving as additional data for assistance. Additionally, in recent years, transformers have achieved outstanding performance in visual tasks, and their large, effective receptive field can provide more discriminative cues for shadow regions. For these reasons, this study proposes a novel semantic segmentation model called MixImages, which can combine polarization data for pixel-level perception. We conducted comprehensive experiments on a polarization dataset of urban scenes. The results showed that the proposed MixImages can achieve an accuracy advantage of 3.43% over the control group model using only RGB images in the unimodal benchmark while gaining a performance improvement of 4.29% in the multimodal benchmark. Additionally, to provide a reference for specific downstream tasks, we also tested the impact of different combinations of polarization types on the overall segmentation accuracy. The proposed MixImages can be a new option for conducting urban scene perception tasks.

SwinDefNet: A Novel Surface Water Mapping Model in Mountain and Cloudy Regions Based on Sentinel-2 Imagery

Surface water plays a pivotal role in the context of climate change, human activities, and ecosystems, underscoring the significance of precise monitoring and observation of surface water bodies. However, the intricate and diverse nature of surface water distribution poses substantial challenges to accurate mapping. The extraction of water bodies from medium-resolution satellite remote sensing images using CNN methods is constrained by limitations in receptive fields and inadequate context modeling capabilities, resulting in the loss of boundary details of water bodies and suboptimal fusion of multi-scale features. The existing research on this issue is limited, necessitating the exploration of novel deep-learning network combinations to overcome these challenges. This study introduces a novel deep learning network combination, SwinDefNet, which integrates deformable convolution and Swin Transformer for the first time. By enhancing the effective receptive field and integrating global semantic information, the model can effectively capture the diverse features of water bodies at various scales, thereby enhancing the accuracy and completeness of water extraction. The model was evaluated on Sentinel-2 satellite images, achieving an overall accuracy of 97.89%, an F1 score of 92.33%, and, notably, an accuracy of 98.03% in mountainous regions. These findings highlight the promising potential of this combined approach for precise water extraction tasks.

Efficient analysis of deep neural networks for vision via biologically-inspired receptive field angles: An in-depth survey

Efficient feature extraction is a pivotal requirement for Deep Neural Network (DNN) models, particularly in the realm of visual tasks where effective feature extraction relies on well-designed receptive fields. Nevertheless, the evaluation and fulfillment of the criteria for “effective receptive fields”, along with their essential conditions, continue to pose numerous unresolved mysteries. In this Analysis we amalgamated principles from the field of biologically inspired vision to reevaluate the role of receptive fields within neural networks, and have proposed a comprehensive design theory aimed at characterizing and guiding the research of efficient DNNs. Our design theory draws from both the principles of biological vision and the latest advances in DNN design theory. It encompasses five fundamental axes of receptive field design research: the Global Artificial Neuron Workspace (GANW), Layered Processing Theory (LPT), Multi-Receptive Field Integration Design (M-RFID), Heterogeneous Receptive Field (HRF), and Receptive Field Plasticity Theory (RFPT). We have applied these design theories to validate traditional and cutting-edge neural networks, conducting in-depth analyses of the outcomes to provide a summary of the current landscape of research on visual receptive fields. Furthermore, we have classified current visual receptive fields into three categories: sample-specific local receptive fields, global spatial receptive fields, and adaptive dynamic receptive fields. From the perspective of perceptual fields and network lightweight design, we have introduced some classic and state-of-the-art DNN models, with a particular emphasis on those employing innovative algorithms that have yielded the most recent achievements. We have concluded by highlighting the key areas of interest in the field of visual receptive fields for future research endeavors.

DILA: Dynamic Gaussian Distribution Fitting and Imitation Learning-Based Label Assignment for tiny object detection

With the rapid advancement of deep learning technology, significant achievements have been made in the field of general object detection. However, challenges still remain in the detection of tiny objects. There are two main drawbacks: (1) static prior information makes the localization of tiny objects relatively fixed; (2) Intersection over Union (IoU) is highly sensitive to deviations in tiny objects. To this end, we propose a Dynamic Gaussian Distribution Fitting and Imitation Learning-Based Label Assignment (DILA) strategy for Tiny Object Detection. Specifically, to address the positional deviation of effective receptive fields at different network layers during Gaussian modeling, DILA first designs an Adaptive Dynamic Calculation Strategy (ADCS) to compute estimation factors for the effective receptive field in different feature spaces, dynamically modeling prior information using Gaussian distribution. Then, DILA introduces a new Balance of Gaussian Scaling-aware Metric (BGSM) to measure the similarity between tiny bounding boxes and predefined anchors, instead of using IoU, which is highly sensitive to tiny pixel shifts, for sample assignment, thereby providing a more accurate basis for label assignment. Finally, a Detail Information Imitation Compensation Module (DIM) is presented to improve and compensate for the detailed information of tiny objects that troubles label assignment, achieving balanced learning for tiny objects. The proposed DILA strategy can be seamlessly integrated into various anchor-based detectors. Extensive experiments were conducted on three publicly available datasets for tiny object detection. The results indicate that when DILA is embedded into Faster RCNN, it outperforms other state-of-the-art methods in terms of detection performance for tiny objects, achieving an improvement in average precision of 10.3%, 1.5%, and 4.2% on AI-TOD, SODA-D, and VisDrone2019, respectively. The source codes and results are available at: https://github.com/chnu-cpl/DILA.

LERFNet: an enlarged effective receptive field backbone network for enhancing visual drone detection

AbstractRecently, the world has witnessed a great increase in drone applications and missions. Drones must be detected quickly, effectively, and precisely when they are being handled illegally. Vision-based anti-drone systems provide an efficient performance compared to radar- and acoustic-based systems. The effectiveness of drone detection is affected by a number of issues, including the drone’s small size, conflicts with other objects, and noisy backgrounds. This paper employs enlarging the effective receptive field (ERF) of feature maps generated from the YOLOv6 backbone. First, RepLKNet is used as the backbone of YOLOv6, which deploys large kernels with depth-wise convolution. Then, to get beyond RepLKNet’s large inference time, a novel LERFNet is implemented. LERFNet uses dilated convolution in addition to large kernels to enlarge the ERF and overcome each other’s problems. The linear spatial-channel attention module (LAM) is used to give more attention to the most informative pixels and high feature channels. LERFNet produces output feature maps with a large ERF and high shape bias to enhance the detection of various drone sizes in complex scenes. The RepLKNet and LERFNet backbones for Tiny-YOLOv6, Tiny-YOLOv6, YOLOv5s, and Tiny-YOLOv7 are compared. In comparison to the aforementioned techniques, the suggested model’s results show a greater balance between accuracy and speed. LERFNet increases the MAP by $$2.8\%$$ 2.8 % , while significantly reducing the GFLOPs and parameter numbers when compared to the original backbone of YOLOv6.

Receptive Field Space for Point Cloud Analysis.

Similar to convolutional neural networks for image processing, existing analysis methods for 3D point clouds often require the designation of a local neighborhood to describe the local features of the point cloud. This local neighborhood is typically manually specified, which makes it impossible for the network to dynamically adjust the receptive field's range. If the range is too large, it tends to overlook local details, and if it is too small, it cannot establish global dependencies. To address this issue, we introduce in this paper a new concept: receptive field space (RFS). With a minor computational cost, we extract features from multiple consecutive receptive field ranges to form this new receptive field space. On this basis, we further propose a receptive field space attention mechanism, enabling the network to adaptively select the most effective receptive field range from RFS, thus equipping the network with the ability to adjust granularity adaptively. Our approach achieved state-of-the-art performance in both point cloud classification, with an overall accuracy (OA) of 94.2%, and part segmentation, achieving an mIoU of 86.0%, demonstrating the effectiveness of our method.

GaitRA: triple-branch multimodal gait recognition with larger effective receptive fields and mixed attention

GaitRA: triple-branch multimodal gait recognition with larger effective receptive fields and mixed attention

Wide-field, high-resolution reconstruction in computational multi-aperture miniscope using a Fourier neural network

Traditional fluorescence microscopy is constrained by inherent trade-offs among resolution, field of view, and system complexity. To navigate these challenges, we introduce a simple and low-cost computational multi-aperture miniature microscope, utilizing a microlens array for single-shot wide-field, high-resolution imaging. Addressing the challenges posed by extensive view multiplexing and non-local, shift-variant aberrations in this device, we present SV-FourierNet, a multi-channel Fourier neural network. SV-FourierNet facilitates high-resolution image reconstruction across the entire imaging field through its learned global receptive field. We establish a close relationship between the physical spatially varying point-spread functions and the network’s learned effective receptive field. This ensures that SV-FourierNet has effectively encapsulated the spatially varying aberrations in our system and learned a physically meaningful function for image reconstruction. Training of SV-FourierNet is conducted entirely on a physics-based simulator. We showcase wide-field, high-resolution video reconstructions on colonies of freely moving C. elegans and imaging of a mouse brain section. Our computational multi-aperture miniature microscope, augmented with SV-FourierNet, represents a major advancement in computational microscopy and may find broad applications in biomedical research and other fields requiring compact microscopy solutions.

Pure large kernel convolutional neural network transformer for medical image registration

Deformable medical image registration is a fundamental and critical task in medical image analysis. Recently, deep learning-based methods have rapidly developed and have shown impressive results in deformable image registration. However, existing approaches still suffer from limitations in registration accuracy or generalization performance. To address these challenges, in this paper, we propose a pure convolutional neural network module (CVTF) to implement hierarchical transformers and enhance the registration performance of medical images. CVTF has a larger convolutional kernel, providing a larger global effective receptive field, which can improve the network’s ability to capture long-range dependencies. In addition, we introduce the spatial interaction attention (SIA) module to compute the interrelationship between the target feature pixel points and all other points in the feature map. This helps to improve the semantic understanding of the model by emphasizing important features and suppressing irrelevant ones. Based on the proposed CVTF and SIA, we construct a novel registration framework named PCTNet. We applied PCTNet to generate displacement fields and register medical images, and we conducted extensive experiments and validation on two public datasets, OASIS and LPBA40. The experimental results demonstrate the effectiveness and generality of our method, showing significant improvements in registration accuracy and generalization performance compared to existing methods. Our code has been available at https://github.com/fz852/PCTNet.

Sparsity-Robust Feature Fusion for Vulnerable Road-User Detection with 4D Radar

Detecting vulnerable road users is a major challenge for autonomous vehicles due to their small size. Various sensor modalities have been investigated, including mono or stereo cameras and 3D LiDAR sensors, which are limited by environmental conditions and hardware costs. Radar sensors are a low-cost and robust option, with high-resolution 4D radar sensors being suitable for advanced detection tasks. However, they involve challenges such as few and irregularly distributed measurement points and disturbing artifacts. Learning-based approaches utilizing pillar-based networks show potential in overcoming these challenges. However, the severe sparsity of radar data makes detecting small objects with only a few points difficult. We extend a pillar network with our novel Sparsity-Robust Feature Fusion (SRFF) neck, which combines high- and low-level multi-resolution features through a lightweight attention mechanism. While low-level features aid in better localization, high-level features allow for better classification. As sparse input data are propagated through a network, the increasing effective receptive field leads to feature maps of different sparsities. The combination of features with different sparsities improves the robustness of the network for classes with few points.

Multiple object detection and tracking from drone videos based on GM-YOLO and multi-tracker

Multiple object tracking in drone videos is a vital vision task with broad application prospects, but most trackers use spatial or appearance clues alone to correlate detections. Our proposed Multi-Tracker uses a novel similarity measure that combines position and appearance information. We designed the GM-YOLO network to provide high-quality detections as input to Multi-Tracker. Add a Coordinate Attention mechanism and a weighted Bidirectional Feature Pyramid Network structure to the Backbone, each feature point's effective receptive field is modeled as a Gaussian distribution. To accurately obtain the motion and appearance features of the object, the adaptive noise covariance Kalman filter is used to get the position information, MB-OSNet network is designed to use global features to learn contour information to retrieve images from a wider field of view while incorporating Part-Level elements that contain more fine-grained data. Finally, the motion and appearance features are jointly compared to realize multi object tracking. The performance of the GM-YOLO object detector and the Multi-Tracker was verified on the VisDrone MOT and UAVDT datasets.

CDYL for infrared and visible light image dense small object detection

To address the phenomenon of many small and hard-to-detect objects in infrared and visible light images, we propose an object detection algorithm CDYL (Convolution to Fully Connect-ed-Deformable Convolution You only Look once) based on the CFC-DC (Convolution to Fully Connected-Deformable Convolution) module. The core operator of CDYL is CFC-DC, making our model not only have a large effective receptive field in infrared and visible light images, but also have adaptive spatial aggregation conditioned by input and task information. As a result, the CDYL reduces the strict inductive bias of traditional CNNs and has long-range dependence for large kernel convolution as well as adaptive spatial aggregation, deeply mining of edge and correlation information in images to enhance sensitivity to small objects, thereby improving performance in dense small object detection tasks. In order to improve the ability of the CFC-DC module to perceive the detailed information of the image, we use the Mish activation function, which has a wider minima which improves the generalization. The effectiveness as well as the generalization of CDYL is evaluated on an infrared image dataset and an UAV image dataset, and it is compared with other state-of-the-art object detection algorithms. Compared to the baseline network YOLOv8l, our model achieved a 3.0% improvement in mAP0.5 in infrared image detection tasks and a 1.1% improvement in mAP0.5 in visible light image detection tasks. The experimental results show that the proposed algorithm achieves superior average precision values on both infrared and visible light images, while maintaining a light weight. Code is publicly available at https://github.com/yangzhu1/CDYL.

MPFANet: a multipath feature aggregation network for steel surface defect detection

As a crucial component of product quality surveillance in the industrial manufacturing field, surface defect detection plays a pivotal role in achieving industrial automation. Although numerous defect detection methods have been widely used in steel production, complicated defect characteristics still severely affect the accuracy and efficiency of defect detection methods. To solve this problem, we propose a multi-path feature aggregation network that can significantly improve steel defect detection performance. Firstly, we design an efficient multi-path aggregation feature extraction module combining gradient path with data path propagation design paradigm as the backbone feature extraction unit based on YOLOv5. Secondly, in the branch path of the backbone, deformable convolution is introduced to expand the effective receptive field and enhance the capability of extracting irregularly distributed defect features. Furthermore, we propose a chained feature enhanced pooling module that refines feature information to obtain feature maps containing more complicated defect details in the pooling layer. Experimental results on the publicly available steel surface defect datasets show that our proposed model achieves a mean average precision of 81.9 % on Northeastern University-DET and 72.1 % on GC10-DET improving by 4.6 % and 3.4 % over the baseline, respectively. The proposed defect detection network can achieve a balance between accuracy and efficiency, and meet the demand in practical production. The detection speed of our work reaches 41.1 frames per second (FPS) and 40.0 FPS on the above two datasets, which demonstrates that our method has achieved comprehensive performance in steel surface defect detection.

ICNet: A Dual-Branch Instance Segmentation Network for High-Precision Pig Counting

A clear understanding of the number of pigs plays a crucial role in breeding management. Computer vision technology possesses several advantages, as it is harmless and labour-saving compared to traditional counting methods. Nevertheless, the existing methods still face some challenges, such as: (1) the lack of a substantial high-precision pig-counting dataset; (2) creating a dataset for instance segmentation can be time-consuming and labor-intensive; (3) interactive occlusion and overlapping always lead to incorrect recognition of pigs; (4) existing methods for counting such as object detection have limited accuracy. To address the issues of dataset scarcity and labor-intensive manual labeling, we make a semi-auto instance labeling tool (SAI) to help us to produce a high-precision pig counting dataset named Count1200 including 1220 images and 25,762 instances. The speed at which we make labels far exceeds the speed of manual annotation. A concise and efficient instance segmentation model built upon several novel modules, referred to as the Instances Counting Network (ICNet), is proposed in this paper for pig counting. ICNet is a dual-branch model ingeniously formed of a combination of several layers, which is named the Parallel Deformable Convolutions Layer (PDCL), which is trained from scratch and primarily composed of a couple of parallel deformable convolution blocks (PDCBs). We effectively leverage the characteristic of modeling long-range sequences to build our basic block and compute layer. Along with the benefits of a large effective receptive field, PDCL achieves a better performance for multi-scale objects. In the trade-off between computational resources and performance, ICNet demonstrates excellent performance and surpasses other models in Count1200, AP of 71.4% and AP50 of 95.7% are obtained in our experiments. This work provides inspiration for the rapid creation of high-precision datasets and proposes an accurate approach to pig counting.