Channels Of Feature Maps Research Articles

Improved Real-Time Detection Transformer-Based Rail Fastener Defect Detection Algorithm

To address the issues of the Real-Time DEtection TRansformer (RT-DETR) object detection model, including poor defect feature extraction in the task of rail fastener defect detection, inefficient use of computational resources, and suboptimal channel attention in the self-attention mechanism, the following improvements were made. Firstly, a Super-Resolution Convolutional Module (SRConv) was designed as a separate component and integrated into the Backbone network, which enhances the image details and clarity while preserving the original image structure and semantic content. This integration improves the model’s ability to extract defect features. Secondly, a channel attention mechanism was integrated into the self-attention module of RT-DETR to enhance the focus on feature map channels, addressing the problem of sparse attention maps caused by the lack of channel attention while saving computational resources. Finally, the experimental results show that compared to the original model, the improved RT-DETR-based rail fastener defect detection algorithm, with an additional 0.4 MB of parameters, achieved a higher accuracy, with a 2.8 percentage point increase in the Mean Average Precision (mAP) across IoU thresholds from 0.5 to 0.9 and a 1.7 percentage point increase in the Average Recall (AR) across the same thresholds.

Improved YOLOv8n for Lightweight Ship Detection

Automatic ship detection is a crucial task within the domain of maritime transportation management. With the progressive success of convolutional neural networks (CNNs), a number of advanced CNN models have been presented in order to detect ships. Although these detection models have achieved marked performance, several undesired results may occur under complex maritime conditions, such as missed detections, false positives, and low detection accuracy. Moreover, the existing detection models endure large number of parameters and heavy computation cost. To deal with these problems, we suggest a lightweight ship model of detection called DSSM–LightNet based upon the improved YOLOv8n. First, we introduce a lightweight Dual Convolutional (DualConv) into the model to lower both the number of parameters and the computational complexity. The principle is that DualConv combines two types of convolution kernels, 3x3 and 1x1, and utilizes group convolution techniques to effectively reduce computational costs while processing the same input feature map channels. Second, we propose a Slim-neck structure in the neck network, which introduces GSConv and VoVGSCSP modules to construct an efficient feature-fusion layer. This fusion strategy helps the model better capture the features of targets of different sizes. Meanwhile, a spatially enhanced attention module (SEAM) is leveraged to integrate with a Feature Pyramid Network (FPN) and the Slim-neck to achieve simple yet effective feature extraction, minimizing information loss during feature fusion. CIoU may not accurately reflect the relative positional relationship between bounding boxes in some complex scenarios. In contrast, MPDIoU can provide more accurate positional information in bounding-box regression by directly minimizing point distance and considering comprehensive loss. Therefore, we utilize the minimum point distance IoU (MPDIoU) rather than the Complete Intersection over Union (CIoU) Loss to further enhance the detection precision of the suggested model. Comprehensive tests carried out on the publicly accessible SeaShips dataset have demonstrated that our model greatly exceeds other algorithms in relation to their detection accuracy and efficiency, while reserving its lightweight nature.

Evaluation of Fracturing Effect of Tight Reservoirs Based on Deep Learning.

The utilization of hydraulic fracturing technology is indispensable for unlocking the potential of tight oil and gas reservoirs. Understanding and accurately evaluating the impact of fracturing is pivotal in maximizing oil and gas production and optimizing wellbore performance. Currently, evaluation methods based on acoustic logging, such as orthogonal dipole anisotropy and radial tomography imaging, are widely used. However, when the fractures generated by hydraulic fracturing form a network-like pattern, orthogonal dipole anisotropy fails to accurately assess the fracturing effects. Radial tomography imaging can address this issue, but it is challenged by high manpower and time costs. This study aims to develop a more efficient and accurate method for evaluating fracturing effects in tight reservoirs using deep learning techniques. Specifically, the method utilizes dipole array acoustic logging curves recorded before and after fracturing. Manual labeling was conducted by integrating logging data interpretation results. An improved WGAN-GP was employed to generate adversarial samples for data augmentation, and fracturing effect evaluation was implemented using SE-ResNet, ResNet, and DenseNet. The experimental results demonstrated that ResNet with residual connections is more suitable for the dataset in this study, achieving higher accuracy in fracturing effect evaluation. The inclusion of the SE module further enhanced model accuracy by adaptively adjusting the weights of feature map channels, with the highest accuracy reaching 99.75%.

A Study of Improved Two-Stage Dual-Conv Coordinate Attention Model for Sound Event Detection and Localization.

Sound Event Detection and Localization (SELD) is a comprehensive task that aims to solve the subtasks of Sound Event Detection (SED) and Sound Source Localization (SSL) simultaneously. The task of SELD lies in the need to solve both sound recognition and spatial localization problems, and different categories of sound events may overlap in time and space, making it more difficult for the model to distinguish between different events occurring at the same time and to locate the sound source. In this study, the Dual-conv Coordinate Attention Module (DCAM) combines dual convolutional blocks and Coordinate Attention, and based on this, the network architecture based on the two-stage strategy is improved to form the SELD-oriented Two-Stage Dual-conv Coordinate Attention Model (TDCAM) for SELD. TDCAM draws on the concepts of Visual Geometry Group (VGG) networks and Coordinate Attention to effectively capture critical local information by focusing on the coordinate space information of the feature map and dealing with the relationship between the feature map channels to enhance the feature selection capability of the model. To address the limitation of a single-layer Bi-directional Gated Recurrent Unit (Bi-GRU) in the two-stage network in terms of timing processing, we add to the structure of the two-layer Bi-GRU and introduce the data enhancement techniques of the frequency mask and time mask to improve the modeling and generalization ability of the model for timing features. Through experimental validation on the TAU Spatial Sound Events 2019 development dataset, our approach significantly improves the performance of SELD compared to the two-stage network baseline model. Furthermore, the effectiveness of DCAM and the two-layer Bi-GRU structure is confirmed by performing ablation experiments.

Convolutional Neural Network‐Based CT Image Segmentation of Kidney Tumours

ABSTRACTKidney tumours are one of the most common tumours in humans and the main current treatment is surgical removal. The CT images are usually manually segmented by a specialist for pre‐operative planning, but this can be influenced by the surgeon's experience and skill and can be time‐consuming. Due to the complex lesions and different morphologies of kidney tumours that make segmentation difficult, this article proposes a convolutional neural network‐based automatic segmentation method for CT images of kidney tumours to address the most common problems of boundary blurring and false positives in tumour segmentation images. The method is highly accurate and reliable, and is used to assist doctors in surgical planning as well as diagnostic treatment, relieving medical pressure to a certain extent. The EfficientNetV2‐UNet segmentation model proposed in this article includes three main parts: feature extractor, reconstruction network and Bayesian decision algorithm. Firstly, for the phenomenon of tumour false positives, the EfficientNetV2 feature extractor, which has high training accuracy and efficiency, is selected as the backbone network, which extracts shallow features such as tumour location, morphology and texture in the CT image by downsampling. Secondly, on the basis of the backbone network, the reconstruction network is designed, which mainly consists of conversion block, deconvolution block, convolution block and output block. Then, the up‐sampling architecture is constructed to gradually recover the spatial resolution of the feature map, fully identify the contextual information and form a complete encoding–decoding structure. Multi‐scale feature fusion is achieved by superimposing all levels of feature map channels on the left and right sides of the network, preventing the loss of details and performing accurate tumour segmentation. Finally, a Bayesian decision algorithm is designed for the edge blurring phenomenon of segmented tumours and cascaded over the output of the reconstruction network, combining the edge features of the original CT image and the segmented image for probability estimation, which is used to improve the accuracy of the model edge segmentation. Medical images in NII special format were converted to Numpy matrix format using python, and then more than 2000 CT images containing only kidney tumours were selected from the KiTS19 dataset as the dataset for the model, and the dimensions were standardised to 128 × 128, and the experimental results show that the model outperforms many other advanced models with good segmentation performance.

A Novel Deep Learning Network Model for Extracting Lake Water Bodies from Remote Sensing Images

Extraction of lake water bodies from remote sensing images provides reliable data support for water resource management, environmental protection, natural disaster early warning, and scientific research, and helps to promote sustainable development, protect the ecological environment and human health. With reference to the classical encoding-decoding semantic segmentation network, we propose the network model R50A3-LWBENet for lake water body extraction from remote sensing images based on ResNet50 and three attention mechanisms. R50A3-LWBENet model uses ResNet50 for feature extraction, also known as encoding, and squeeze and excitation (SE) block is added to the residual module, which highlights the deeper features of the water body part of the feature map during the down-sampling process, and also takes into account the importance of the feature map channels, which can better capture the multiscale relationship between pixels. After the feature extraction is completed, the convolutional block attention module (CBAM) is added to give the model a global adaptive perception capability and pay more attention to the water body part of the image. The feature map is up-sampled using bilinear interpolation, and the features at different levels are fused, a process also known as decoding, to finalize the extraction of the lake water body. Compared with U-Net, AU-Net, RU-Net, ARU-Net, SER34AUNet, and MU-Net, the R50A3-LWBENet model has the fastest convergence speed and the highest MIoU accuracy with a value of 97.6%, which is able to better combine global and local information, refine the edge contours of the lake’s water body, and have stronger feature extraction capability and segmentation performance.

A Lightweight SAR Image Ship Detection Method Based on Improved Convolution and YOLOv7

The airborne and satellite-based synthetic aperture radar enables the acquisition of high-resolution SAR oceanographic images in which even the outlines of ships can be identified. The detection of ship targets from SAR images has a wide range of applications. Due to the density of ships in SAR images, the extreme imbalance between foreground and background clutter, and the diversity of target sizes, achieving lightweight and highly accurate multi-scale ship target detection remains a great challenge. To this end, this paper proposed an attention mechanism for multi-scale receptive fields convolution block (AMMRF). AMMRF not only makes full use of the location information of the feature map to accurately capture the regions in the feature map that are useful for detection results, but also effectively captures the relationship between the feature map channels, so as to better learn the relationship between the ship and the background. Based on this, a new YOLOv7-based ship target detection method, You Only Look Once SAR Ship Identification (YOLO-SARSI), was proposed, which acquires the abstract semantic information extracted from the high-level convolution while retaining the detailed semantic information extracted from the low-level convolution. Compared to the deep learning detection methods proposed by previous authors, our method is more lightweight, only 18.43 M. We examined the effectiveness of our method on two SAR image public datasets: the High-Resolution SAR Images Dataset (HRSID) and the Large-Scale SAR Ship Detection Dataset-v1.0 (LS-SSDD-V1.0). The results show that the average accuracy (AP50) of the detection method YOLO-SARSI proposed in this paper on the HRSID and LS-SSDD-V1.0 datasets is 2.6% and 3.9% higher than that of YOLOv7, respectively.

YOLO-based marine organism detection using two-terminal attention mechanism and difficult-sample resampling

The presence of various types of noise in images of marine-life datasets, as well as the class imbalances in underwater datasets, can exacerbate the difficulty in achieving effective object detection. To address this problem, we proposed you only look once (YOLO)-based marine organism detection using a two-terminal attention mechanism and difficult-sample resampling process. First, a residual building unit (RBU) module with a two-terminal attention mechanism (RBU-TA) was proposed, incorporating a reinforced channel attention mechanism into a shortcut of the residual structure. The proposed method adaptively compressed noisy feature map channels, providing rich shallow image information for high-level deep convolutional features while avoiding shallow noise pollution. To address the imbalance of marine biological image classes, difficult-sample resampling was combined with a focal loss function to suppress excessive background negative samples and retrain targets that could be difficult to distinguish, thus improving their detection accuracy. Finally, the proposed method was validated using the underwater robot professional competition (URPC) and real-world underwater object detection (RUOD) datasets, and the mean average precision (MAP) values of the results improved by 10% and 7%, respectively. The proposed method greatly improved the target detection accuracy of organisms in complex marine environments.

A novel full-convolution UNet-transformer for medical image segmentation

The Transformer-based methods are still unable to effectively model local contexts although they make up for the deficiency of remote information dependencies for approaches based on small kernel CNNs. To overcome such a shortage, this paper proposes a novel full-convolution UNet Transformer model, FC-UNETTR, for medical image segmentation. First, a novel global–local attention module is proposed by utilizing multiple small kernels of different sizes for depth-wise convolutions to expand the receptive field of the network model, increase the remote dependence of semantic information in the encoder stage, and also improve the feature extraction capability of the network for fuzzy edges. Then, a reparametrized feedforward network is developed to further improve the local information extraction and mitigate the coupling between feature maps such that the relationship between feature map channels can be better revealed. Furthermore, the skip connection and decoder are redesigned by constructing a dense multiscale module instead of traditional ResNet modules to mitigate semantic bias. Benefiting from the above improvements, the constructed FC-UNETTR without pre-training demonstrates a strong capability to extract local features and capture long-range dependencies of images in medical image segmentation. Experiments show that FC-UNETTR achieves an excellent performance of 85.67% for DSC and 7.82 for HD metrics on the Synapse dataset with fewer model parameters compared with state-of-the-art networks. Furthermore, DSC reaches 92.46% and 94.76% on the ACDC dataset and the private dataset of oral graft bone, respectively, outperforming some of the latest medical image segmentation models..

SSD with multi-scale feature fusion and attention mechanism

In the field of the Internet of Things, image acquisition equipment is the very important equipment, which will generate lots of invalid data during real-time monitoring. Analyzing the data collected directly from the terminal by edge calculation, we can remove invalid frames and improve the accuracy of system detection. SSD algorithm has a relatively light and fast detection speed. However, SSD algorithm do not take full advantage of both shallow and deep information of data. So a multiscale feature fusion attention mechanism structure based on SSD algorithm has been proposed in this paper, which combines multiscale feature fusion and attention mechanism. The adjacent feature layers for each detection layer are fused to improve the feature information expression ability. Then, the attention mechanism is added to increase the attention of the feature map channels. The results of the experiments show that the detection accuracy of the optimized model is improved, and the reliability of edge calculation has been improved.

DualConv: Dual Convolutional Kernels for Lightweight Deep Neural Networks.

Convolutional neural network (CNN) architectures are generally heavy on memory and computational requirements which make them infeasible for embedded systems with limited hardware resources. We propose dual convolutional kernels (DualConv) for constructing lightweight deep neural networks. DualConv combines 3×3 and 1×1 convolutional kernels to process the same input feature map channels simultaneously and exploits the group convolution technique to efficiently arrange convolutional filters. DualConv can be employed in any CNN model such as VGG-16 and ResNet-50 for image classification, you only look once (YOLO) and R-CNN for object detection, or fully convolutional network (FCN) for semantic segmentation. In this work, we extensively test DualConv for classification since these network architectures form the backbone for many other tasks. We also test DualConv for image detection on YOLO-V3. Experimental results show that, combined with our structural innovations, DualConv significantly reduces the computational cost and number of parameters of deep neural networks while surprisingly achieving slightly higher accuracy than the original models in some cases. We use DualConv to further reduce the number of parameters of the lightweight MobileNetV2 by 54% with only 0.68% drop in accuracy on CIFAR-100 dataset. When the number of parameters is not an issue, DualConv increases the accuracy of MobileNetV1 by 4.11% on the same dataset. Furthermore, DualConv significantly improves the YOLO-V3 object detection speed and improves its accuracy by 4.4% on PASCAL visual object classes (VOC) dataset.

FFENet: frequency-spatial feature enhancement network for clothing classification.

Clothing analysis has garnered significant attention, and within this field, clothing classification plays a vital role as one of the fundamental technologies. Due to the inherent complexity of clothing scenes in real-world environments, the learning of clothing features in such complex scenes often encounters interference. Because clothing classification relies on the contour and texture information of clothing, clothing classification in real scenes may lead to poor classification results. Therefore, this paper proposes a clothing classification network based on frequency-spatial domain conversion. The proposed network combines frequency domain information with spatial information and does not compress channels. It aims to enhance the extraction of clothing features and improve the accuracy of clothing classification. In our work, (1) we combine the frequency domain information and spatial information to establish a clothing feature extraction clothing classification network without compressed feature map channels, (2) we use the frequency domain feature enhancement module to realize the preliminary extraction of clothing features, and (3) we introduce a clothing dataset in complex scenes (Clothing-8). Our network achieves a top-1 model accuracy of 93.4% on the Clothing-8 dataset and 94.62% on the Fashion-MNIST dataset. Additionally, it also achieves the best results in terms of top-3 and top-5 metrics on the DeepFashion dataset.

ResNet Based on Multi-Feature Attention Mechanism for Sound Classification in Noisy Environments

Environmental noise affects people’s lives and poses challenges for urban sound classification. Traditional algorithms such as Mel frequency cepstral coefficients (MFCCs) struggle due to audio signal complexity. This study applied an attention mechanism to a deep residual network (ResNet) deep learning network to overcome the structural impact of urban noise on audio signals and improve classification accuracy. We propose a three-feature fusion ResNet + attention method (Net50_SE) to maximize information representation in environmental sound signals. This method uses residual structured convolutional neural networks (CNNs) for feature extraction in sound classification tasks. Additionally, an attention module is added to suppress environmental noise impact and focus on different feature map channels. The experimental results demonstrate the effectiveness of our method, achieving 93.2% accuracy compared with 82.87% with CNN and 84.77% with long short-term memory (LSTM). Our model provides higher accuracy and confidence in urban sound classification.

Surface defect detection and classification of steel using an efficient Swin Transformer

Detecting steel-surface defects is a crucial phase in steel manufacturing; however, accurately completing the detection task is challenging. The Swin Transformer, a self-attention-based model, has shown strong performance in the field of computer vision to enhance the adaptability of the Swin Transformer to the task of steel-surface defect detection, and a new network architecture called the LSwin Transformer is proposed in this study. First, in the downsampling process, we propose a convolutional embedding module and an attention patch merging module, which simultaneously strengthen the connections between the feature map channels, reduce the resolution, and increase image information retention. Second, we propose an effective window shift strategy and a convenient computation approach to make a complete defect between patches have more opportunity to obtain interactive computing. Finally, to combine the feature extraction capability of convolutional neural networks with the global dependency building capability of the Swin Transformer, we propose a depth multilayer perceptron module. Numerous experiments were conducted on a steel-surface defect dataset. The results demonstrated that the detection effect of our model outperformed competing methods, with a mean average precision of 81.2 %. In the ablation study, we verified the effectiveness of each module and initialized the parameters of the model through transfer learning to accelerate the convergence of the model. Therefore, the proposed LSwin Transformer has significant potential for detecting steel-surface defects.

Dilated Multi-Temporal Modeling for Action Recognition

Action recognition involves capturing temporal information from video clips where the duration varies with videos for the same action. Due to the diverse scale of temporal context, uniform size kernels utilized in convolutional neural networks (CNNs) limit the capability of multiple-scale temporal modeling. In this paper, we propose a novel dilated multi-temporal (DMT) module that provides a solution for modeling multi-temporal information in action recognition. By using dilated convolutions with different dilation rates in different feature map channels, the DMT module captures information at multiple scales without the need for costly multi-branch networks, input-level frame pyramids, or feature map stacking that previous works have usually incurred. Therefore, this approach enables the integration of temporal information from multiple scales. In addition, the DMT module can be integrated into existing 2D CNNs, making it a straightforward and intuitive solution for addressing the challenge of multi-temporal modeling. Our proposed method has demonstrated promising results in performance and has achieved about 2% and 1% accuracy improvement on FineGym99 and SthV1. We conducted an empirical analysis that demonstrates how DMT improves the classification accuracy for action classes with varying durations.

Reparametric Sampling for Small Object Image Recognition

It is difficult to balance detection speed and detection accuracy in current methods for sampling on feature maps. We combine the parameterless interpolation algorithm used in the traditional method with the algorithm in deep learning to improve shortcomings of the traditional interpolation algorithm by introducing learnable parameters into the algorithm. The new algorithm guarantees the features of the original image by calculating the correlation between feature map channels and pixels to ensure the correct restoration of the image features during up-sampling. In this paper, we propose a new up-sampling operation model — Reparametric Blend Up-sampling (RBU), which uses the method of rearranging pixels to obtain a feature map that matches the up-sampling size, uses blending to compute inter-pixel and inter-channel correlations and back-propagates the model to learn the appropriate parameters to ensure feature extraction. In the field of object detection, this model replaces the original up-sampling operation and improves the mAP mean value by 1.1% and the detection of small target objects by 3.1% in the VOC dataset with the richest application scenarios.

YOLOv5s-SA: Light-Weighted and Improved YOLOv5s for Sperm Detection

Sperm detection performance is particularly critical for sperm motility tracking. However, there are a large number of non-sperm objects, sperm occlusion and poorly detailed texture features in semen images, which directly affect the accuracy of sperm detection. To solve the problem of false detection and missed detection in sperm detection, a multi-sperm target detection model, Yolov5s-SA, with an SA attention mechanism is proposed based on the YOLOv5s algorithm. Firstly, a depthwise, separable convolution structure is used to replace the partial convolution of the backbone network, which can ensure stable precision and reduce the number of model parameters. Secondly, a new multi-scale feature fusion module is designed to enhance the perception of feature information to supplement the positional information and high-resolution of the deep feature map. Finally, the SA attention mechanism is integrated into the neck network before the output of the feature map to enhance the correlation between the feature map channels and improve the fine-grained feature fusion ability of YOLOv5s. Experimental results show that compared with various YOLO algorithms, the proposed algorithm improves the detection accuracy and speed to a certain extent. Compared with the YOLOv3, YOLOv3-spp, YOLOv5s and YOLOv5m models, the average accuracy increases by 18.1%, 15.2%, 6.9% and 1.9%, respectively. It can effectively reduce the missed detection of occluded sperm and achieve lightweight and efficient multi-sperm target detection.

SAR Image Classification Using Gated Channel Attention Based Convolutional Neural Network

Algorithms combining CNN (Convolutional Neural Network) and super-pixel based smoothing have been proposed in recent years for Synthetic Aperture Radar (SAR) image classification. However, the smoothing may lead to the damage of details. To solve this problem the feature fusion strategy is utilized, and a novel adaptive fusion module named Gated Channel Attention (GCA) is designed in this paper. In this module, the relevance between channels is embedded into the conventional gated attention module to emphasize the variation in contribution on classification results between channels of feature-maps, which is not well considered by the conventional gated attention module. A GCA-CNN network is then constructed for SAR image classification. In this network, feature-maps corresponding to the original image and the smoothed image are extracted, respectively, by feature-extraction layers and adaptively fused. The fused features are used to obtain the results. Classification can be performed by the GCA-CNN in an end-to-end way. By the adaptive feature fusion in GCA-CNN, the smoothing of misclassification and the detail keeping can be realized at the same time. Experiments have been performed on one elaborately designed synthetic image and three real world SAR images. The superiority of the GCA-CNN is demonstrated by comparing with the conventional algorithms and the relative state-of-the-art algorithms.

Generalized Ridge Regression-Based Channelwise Feature Map Weighted Reconstruction Network for Fine-Grained Few-Shot Ship Classification

Fine-grained ship classification (FGSCR) has many applications in military and civilian fields. In recent years, deep learning has been widely used for classification tasks, and its success is inseparable from that of big data. However, ship images are valuable, with only a few images of a specific category being obtained, leading to the fine-grained few-shot ship classification problem. In addition, feature map channels contain distinct characteristics and discriminative details, which significantly influence FGSCR. Intuitively, channels with distinct characteristics should be assigned larger weights for classification, but most few-shot learning methods treat the channels equally. Therefore, we propose a generalized ridge-regression-based channelwise feature map weighted reconstruction network to address these issues. First, we reconstruct the query feature map by assigning different weights to the support feature map channels using the generalized ridge regression method. The channels with large discriminative details contribute more toward reconstruction. Second, we propose a support channel weight module to calculate the channel weight matrix used in the generalized ridge regression method. Finally, based on the reconstructed query feature map, we can calculate the reconstruction error. The reconstruction error is adopted as the distance metric. Our proposed method achieves excellent performance on the fine-grained ship, bird, aircraft, and WHU-RS19 datasets compared with other representative few-shot learning methods. Considering the limited studies on the fine-grained few-shot ship classification problem, we believe that our work is of great significance.

Segmentation of Clinical Target Volume From CT Images for Cervical Cancer Using Deep Learning.

Introduction: Segmentation of clinical target volume (CTV) from CT images is critical for cervical cancer brachytherapy, but this task is time-consuming, laborious, and not reproducible. In this work, we aim to propose an end-to-end model to segment CTV for cervical cancer brachytherapy accurately. Methods: In this paper, an improved M-Net model (Mnet_IM) is proposed to segment CTV of cervical cancer from CT images. An input and an output branch are both proposed to attach to the bottom layer to deal with CTV locating challenges due to its lower contrast than surrounding organs and tissues. A progressive fusion approach is then proposed to recover the prediction results layer by layer to enhance the smoothness of segmentation results. A loss function is defined on each of the multiscale outputs to form a deep supervision mechanism. Numbers of feature map channels that are directly connected to inputs are finally homogenized for each image resolution to reduce feature redundancy and computational burden. Result: Experimental results of the proposed model and some representative models on 5438 image slices from 53 cervical cancer patients demonstrate advantages of the proposed model in terms of segmentation accuracy, such as average surface distance, 95% Hausdorff distance, surface overlap, surface dice, and volumetric dice. Conclusion: A better agreement between the predicted CTV from the proposed model Mnet_IM and manually labeled ground truth is obtained compared to some representative state-of-the-art models.