Banner | Promotion
Cover of Volume 38, Issue 11 of “Liquid Crystal and Display” 2023
This Special Issue PrefaceFrom fullerenes to carbon nanotubes and then to graphene, carbon-based nanomaterials play a significant role in the development of modern science and technology due to their superior physical, mechanical, and chemical properties. Carbon dots, as an emerging zero-dimensional carbon-based material, have attracted great interest since their accidental discovery in 2004 due to their excellent photoluminescence properties and biocompatibility. From the initial blue fluorescent carbon dots to the current carbon dots that cover the entire spectrum of luminescence; from short-lived fluorescent carbon dots to long-lived phosphorescent and chemiluminescent carbon dots. The luminous quantum efficiency has also improved from less than 10% to over 90%, becoming an important branch in the field of photoluminescence. Their excellent luminescent properties and biocompatibility make carbon dots have important application prospects in fields such as luminescent devices, biological imaging, ion detection, fluorescent inks, and information encryption.When it comes to “perovskite”, it inevitably evokes a certain obsession among some “believers” and a bit of madness! Is there a universal deity protecting it? If not, why does it continue to surprise the world in such a short time? From PSC to PeLED, the photoelectric/electro-optical conversion efficiency is rapidly catching up, likely triggering a new revolution in the solar photovoltaic and quantum dot display fields! What exactly is perovskite—a “gem” bestowed by nature or a “hard stone”? How many people have tried to unveil its mysterious veil, and how many have harbored a desire to conquer it? What is so “magical” about perovskite? To conquer it, one must first excavate, reveal, and understand it. The special issue “Perovskite Luminescent/Optoelectronic Materials and Devices” aims to provide a platform for researchers engaged in luminescence and related optoelectronic fields to share their theoretical, technical, and application research findings, hoping you will spread the word and enjoy the feast!
Digital image processing refers to the methods and techniques for removing noise, enhancing, restoring, segmenting, and extracting features from images using computers. With the rapid development of technologies such as computers, artificial intelligence, and deep learning, digital image processing technology has been widely applied in fields such as aerospace, biomedicine, communication engineering, industry and engineering, military public security, cultural arts, and e-commerce. In the future, it will also involve more fields, such as smart cities, intelligent manufacturing, virtual reality, and augmented reality. Many research groups at home and abroad have conducted extensive and in-depth fundamental and applied research in this field.
Against this backdrop, to promote the advancement and application of digital image processing technology, the editorial department of “Liquid Crystal and Display” organized a special issue on “Digital Image Processing” to showcase theoretical, technical, and application research results focusing on target detection and recognition, image enhancement, and image segmentation. This special issue includes 14 outstanding papers from institutions such as the Changchun Institute of Optics, Fine Mechanics and Physics of the Chinese Academy of Sciences, Fuzhou University, Shanghai Ocean University, Shaanxi University of Science and Technology, Guizhou University, and Liaoning University of Technology.
In the area of target detection and recognition, Sun Haijiang and others proposed a method based on attention mechanisms for detecting weak targets in infrared images under complex scenes where the proportion of weak target pixels is low and feature details are not obvious, making it difficult to extract target features and resulting in low detection accuracy. This method can effectively detect weak infrared targets under various complex backgrounds, demonstrating good robustness and adaptability. Guo Jielong and others proposed a new boundary box loss function based on intersection over union (IoU), addressing regression obstacles in special cases and improving the correlation between regression tasks and evaluation metrics while maintaining the integrity of boundary box center point regression attributes, enhancing detection accuracy and convergence speed. Wang Zhi and others proposed a differential confocal tilt measurement sensor, which accurately locates the focal point using the differential response signal obtained from axial scanning and analyzes the field strength distribution of the microscope’s pupil plane to extract the peak position of the light spot image, thus achieving precise tilt measurement. This sensor provides a new method for high-precision contour measurement of freeform surfaces. Qiao Jihong and others proposed an automatic evaluation method that integrates subjective scene imaging color and white balance from the camera, fully extracting relevant features from color images and simulating human visual perception characteristics to evaluate image colors. The proposed method can improve evaluation efficiency, save manpower, and yield evaluation results that are consistent with human subjective judgments. Yang Chen and others proposed a lightweight network model based on the attention mechanism for facial expression recognition, addressing issues of large network models and low accuracy in the ResNet18 network model. This model can achieve facial expression recognition with fewer parameters and higher accuracy. Pan Hao and others proposed an attribute scattering center matching method based on deep belief networks (DBN) to tackle the challenges of SAR target recognition under extended operational conditions, defining similarity measurement criteria based on the constructed attribute scattering center matching relationships and determining the category of test samples based on maximum similarity principles. This method demonstrates good effectiveness and robustness for SAR target recognition. Sun Haijiang and others designed a Winograd algorithm convolutional neural network accelerator based on field-programmable gate arrays (FPGAs), achieving superior convolution layer computing performance and efficiency compared to other FPGA accelerator designs, capable of completing hardware-accelerated computations for remote sensing image classification tasks with high energy efficiency. Zhao Weichao and others proposed a supervised contrastive learning-based classification method for acquiring and analyzing open-source aerospace information, addressing the issues of lengthy sample content and limited related sample numbers. This method effectively utilizes publicly available data resources to extract open-source aerospace information and generate corresponding images, providing significant value for the analysis and research of aerospace information. Wei Xian and others proposed a pose-agnostic replay method for online incremental learning of multi-pose targets, mitigating catastrophic forgetting when facing multi-pose targets in online incremental learning, demonstrating excellent stability and plasticity.
In the area of image enhancement, Yuan Hongchun and others proposed a method that combines a lightweight feature fusion network with multi-color model correction for underwater image enhancement, effectively correcting color casts and improving brightness, saturation, and contrast, resulting in more natural and rich colors in enhanced images. Wang Yuqing and others proposed a neural network non-uniformity correction algorithm improved by side window filtering, which can effectively remove non-uniformity noise from images without obvious ghosting phenomena, demonstrating significant advantages in both non-uniformity correction effects and algorithm running efficiency, providing new research ideas for achieving real-time non-uniformity correction on low-power mobile platforms. Ma Zhenling and others proposed an image mixed distortion correction method based on an improved U-Net network, which transforms the distortion image correction problem into a prediction problem of pixel-by-pixel coordinate changes using deep learning methods, eliminating the complex mathematical model calculations found in traditional methods.
In the area of image segmentation, Xu Yang and others proposed a knowledge distillation-based feature refinement semantic segmentation model, FRKDNet, which can better separate effective content from noise in distilled knowledge, achieving superior segmentation performance and robustness compared to mainstream methods. Zou Yaobin and others proposed a fast two-dimensional cumulative residual Tsallis entropy threshold segmentation method based on the two-dimensional survival function of images, effectively distinguishing targets and backgrounds in images with various histogram patterns.
The series of digital image processing research results presented above aims to provide references and inspiration for a wide range of readers and peers in related fields.
Editorial Department of “Liquid Crystal and Display”
October 16, 2023
This Issue Directory
Target Detection and Recognition
Research on Detection Method for Weak Infrared Targets in Complex Backgrounds Based on Attention Mechanism
Authors:Liu Ying, Sun Haijiang, Zhao Yongxian
Abstract:To address the challenges of weak target pixel proportions being low and feature details being unclear in infrared images under complex scenes, making target feature extraction difficult and detection accuracy low, a method based on attention mechanisms for detecting weak infrared targets in complex backgrounds is proposed. This method is based on the YOLOv5 network, designing a SimAMC3 attention mechanism module to optimize the feature extraction layer of the network; designing a target detection head and changing the depth at which feature extraction begins by adding feature fusion layers to obtain a new weak target detection layer, allowing shallow feature layers to better retain spatial information of weak targets; improving the prediction box selection method to enhance detection accuracy for targets that are close together or overlapping. Experiments selected two SIRST infrared weak target image datasets for labeling and training. Experimental results show that the improved algorithm achieves an average precision mean (mAP) improvement of 4.8% and 7.1% compared to the original YOLOv5 algorithm, effectively detecting weak infrared targets under various complex backgrounds, demonstrating good robustness and adaptability, and can be effectively applied to weak infrared target detection in complex backgrounds.
Keywords:Deep Learning; Weak Infrared Targets; Target Detection; Attention Mechanism
Read the original text (copy or click the link below):
https://cjlcd.lightpublishing.cn/thesisDetails#10.37188/CJLCD.2023-0030
Target Detection Algorithm Based on Adaptive Focusing CRIoU Loss
Authors:Xiao Zhenjiu, Zhao Haoze, Zhang Lili, Xia Yu, Guo Jielong, Yu Hui, Li Chenglong, Wang Liwen
Abstract:In target detection tasks, the traditional boundary box regression loss function has an irrelevant content with the evaluation standard IoU (Intersection over Union), and there are certain unreasonable aspects regarding the regression attributes of the boundary box, leading to incomplete regression attributes, reducing detection accuracy and convergence speed, and even causing regression obstacles. Additionally, there is an imbalance in samples in regression tasks, where a large number of low-quality samples affect loss convergence. To improve detection accuracy and regression convergence speed, a new boundary box regression loss function is proposed. First, the design idea is determined and the paradigm of IoU series loss functions is designed; then, based on IoU loss, a distance penalty term for the center point of the boundary box is introduced, which forms a ratio of the perimeter of the rectangular area formed by two centers and the perimeter of the minimum enclosing rectangle formed by the two boxes, and the improved IoU loss is applied in Non-Maximum Suppression (NMS) processing. Next, the width and height errors of the two boxes and the square of the width and height of the minimum enclosing box are introduced as width and height penalty terms, determining the CRIoU (Complete Relativity IoU) loss function. Finally, an adaptive weighting factor is added to the CRIoU to weigh the regression loss of high-quality samples, defining the Adaptive Focal CRIoU (AF-CRIoU). Experimental results show that using the AF-CRIoU loss function compared to traditional non-IoU series losses results in a maximum relative increase in detection accuracy of 8.52%, and compared to CIoU series losses, a maximum relative increase of 2.69%. Using the A-CRIoU-NMS (Around CRIoU NMS) method compared to the original NMS method results in a detection accuracy improvement of 0.14%. The AF-CRIoU loss is applied to hard hat detection, achieving good detection results.
Keywords:Target Detection; Boundary Box Regression; IoU Loss Function; Non-Maximum Suppression; Adaptive Focal Loss
Read the original text (copy or click the link below):
https://cjlcd.lightpublishing.cn/thesisDetails#10.37188/CJLCD.2023-0005
Inclination Measurement Sensor Based on Differential Confocal
Authors:Wang Tingyu, Wang Zhi, Yang Yongqiang, Mi Xiaotao, Wang Jianli, Yao Kainan, Cheng Xue
Abstract:To solve the problem that traditional laser differential confocal microscopes (LDCM) cannot perform high-precision tilt angle measurements while measuring distance, a differential confocal tilt measurement sensor is proposed. When measuring an inclined surface, this sensor first accurately locates the focal point using the differential response signal obtained from axial scanning, then analyzes the field strength distribution of the microscope’s pupil plane and extracts the peak position of the light spot image, thus achieving precise tilt measurement. First, a light field distribution model after the focused beam reflects off the inclined surface to be measured is established to analyze the field strength distribution of the microscope’s pupil plane at different tilt angles. Then, based on the analysis of the inclined light spot characteristics, a method using an improved Meanshift algorithm to extract the peak position of the light spot is proposed. Finally, experiments validate the effectiveness of the sensor for tilt measurement. Experimental results show that the sensor’s average measurement error for tilt angles (0~8°) is 0.011°, and for tilt directions (0~360°) is 0.128°, meeting the requirements for measuring the tilt angle of the surface to be measured during the detection process using a differential confocal non-contact optical probe. This sensor provides a new method for high-precision contour measurement of freeform surfaces.
Keywords:Non-Contact Optical Probe; Differential Confocal; 3D Detection; Tilt Measurement; Peak Extraction
Read the original text (copy or click the link below):
https://cjlcd.lightpublishing.cn/thesisDetails#10.37188/CJLCD.2023-0045
Color Evaluation of Color Images Based on Color Space
Authors: Zhang Xin, Qiao Jihong, Zhang Huiyan, Zhang Yan, Zhang Xin, Xu Jiping
Abstract:Given the necessity of color evaluation for mobile imaging quality, an automatic evaluation method (CIQA) that integrates subjective scene imaging color and white balance is proposed, fully extracting relevant features from color images and simulating human visual perception characteristics to evaluate image colors. First, a combined method of Scale-invariant feature transform (SIFT) and transmission transformation is used to identify the positions corresponding to the ColorChecker standard 24-color card in the subjective images; then a least squares model with minimum deviation rate is constructed, and expert weighting methods and entropy weighting methods are used to calculate the proportion of weight distribution for color restoration and white balance indicators; finally, an improvement to the TOPSIS (Technique for Order Preference by Similarity to an Ideal Solution) is made based on multi-indicator weight values to determine the closeness of each scheme to typical positive and negative ideal schemes, achieving a ranking of the quality of smartphone imaging colors. Experiments were conducted on images collected from real scenes and compared with two existing decision-making methods for verification. Results show that the proposed method can improve evaluation efficiency, save manpower, and yield evaluation results that are consistent with human subjective judgments.
Keywords:Target Recognition; Indicators; Least Squares of Deviation Rate; Color; Smartphone
Read the original text (copy or click the link below):
https://cjlcd.lightpublishing.cn/thesisDetails#10.37188/CJLCD.2023-0007
Facial Expression Recognition Based on Lightweight ResNet Network with Attention Mechanism
Authors:Zhao Xiao, Yang Chen, Wang Ruonan, Li Yuechen
Abstract:To address issues of large network models and low accuracy in facial expression recognition using the ResNet18 network model, a lightweight network model based on attention mechanisms (Multi-Scale CBAM Lightweight ResNet, MCLResNet) is proposed, capable of achieving facial expression recognition with fewer parameters and higher accuracy. First, ResNet18 is used as the backbone network for feature extraction, and group convolution is introduced to reduce the parameter count of ResNet18; the inverted residual structure increases network depth, optimizing image feature extraction. Second, in the channel attention module of CBAM (Convolutional Block Attention Module), the shared fully connected layer is replaced with a 1×3 convolution module, effectively reducing the loss of channel information; a multi-scale convolution module is added in the spatial attention module of CBAM to obtain spatial feature information at different scales. Finally, the multi-scale spatial feature fusion CBAM module (Multi-Scale CBAM, MSCBAM) is added to the lightweight ResNet model, effectively enhancing the feature expression capability of the network model, and an additional fully connected layer is added to the output layer of the network model to increase the non-linear representation during output. Experimental results on the FER2013 and CK+ datasets show that the proposed model reduces the parameter count by 82.58% compared to ResNet18 while achieving better recognition accuracy.
Keywords: Lightweight ResNet Network; Multi-Scale Spatial Feature Fusion; Facial Expression Recognition; Attention Mechanism
Read the original text (copy or click the link below):
https://cjlcd.lightpublishing.cn/thesisDetails#10.37188/CJLCD.2023-0046
Attribute Scattering Center Matching Based on Deep Belief Network and Its Application in SAR Image Target Recognition
Authors:Xu Yanlong, Pan Hao, Ding Baiyuan
Abstract:Synthetic Aperture Radar (SAR) image target recognition is an important application of SAR image interpretation. To enhance the robustness of SAR target recognition, this paper proposes an attribute scattering center matching method based on Deep Belief Network (DBN). The attribute scattering center parameters are rich in features and can well reflect the local scattering characteristics of the target. DBN leverages the advantages of deep learning to achieve robust matching between test samples and template samples of scattering centers, and can adapt well to noise interference and partial missing conditions. Based on the constructed attribute scattering center matching relationships, similarity measurement criteria are defined. The category of the test sample is determined based on the principle of maximum similarity. Experiments were conducted based on the MSTAR dataset, and validation shows that the proposed method demonstrates good effectiveness and robustness for SAR target recognition.
Keywords:Synthetic Aperture Radar; Target Recognition; Attribute Scattering Center; Deep Belief Network
Read the original text (copy or click the link below):
https://cjlcd.lightpublishing.cn/thesisDetails#10.37188/CJLCD.2023-0052
Design and Implementation of FPGA-Based Winograd Algorithm Convolutional Neural Network Accelerator
Authors:Niuzhao Xu, Sun Haijiang
Abstract:To accelerate the computation of convolutional neural networks in low-power, edge computing scenarios, a Winograd algorithm convolutional neural network accelerator based on field-programmable gate arrays (FPGAs) is designed. First, image data and weight data are quantized to 8-bit fixed-point numbers, and a quantization process in the hardware convolution calculation is designed to enhance data transmission and computation speeds. Next, an input data cache reuse module is designed to merge multi-input channel data for transmission, reusing row-overlapping data. Then, a Winograd pipeline convolution module is designed, achieving column data combination reuse, thereby maximizing the reuse of on-chip data and reducing on-chip data storage occupancy and bandwidth pressure. Finally, the accelerator is deployed on Xilinx’s ZCU104 development board. Experimental validation shows that the convolution layer computation performance of the accelerator reaches 354.5 GOPS, and on-chip DSP computation efficiency reaches 0.69, achieving more than a 1.6 times improvement compared to related research. This accelerator can complete remote sensing image classification tasks with high energy efficiency.
Keywords:Convolutional Neural Networks; Field-Programmable Gate Arrays; Winograd Algorithm; Pipeline; Parallel Computing
Read the original text (copy or click the link below):
https://cjlcd.lightpublishing.cn/thesisDetails#10.37188/CJLCD.2023-0013
Acquisition and Image Generation of Aerospace Information Based on Supervised Contrastive Learning
Authors:Qi Yichen, Zhao Weichao
Abstract:To improve the efficiency of acquiring open-source aerospace information and address issues such as lengthy content and limited quantities of open-source aerospace information, which often lead to poor robustness in commonly used text classification models and less intuitive text information, this paper proposes a supervised contrastive learning-based classification method for aerospace information. This method processes and analyzes open-source information based on a Bidirectional Long Short-Term Memory (BiLSTM) network with an attention mechanism, efficiently filtering aerospace-related information and utilizing the unCLIP (un-Contrastive Language-Image Pre-Training) model to generate corresponding images. Experimental results show that compared to commonly used text classification methods such as CNN (Convolutional Neural Networks), BiLSTM, Transformer, and BiLSTM-Attention, this method performs well in accuracy, recall, and F1-Score, with an F1-Score reaching 0.97, and presents information in image form, making it clearer and more intuitive. This method can effectively utilize publicly available data resources to extract open-source aerospace information and generate corresponding images, providing significant value for the analysis and research of aerospace information.
Keywords:Supervised Text Classification; Contrastive Learning; Text-to-Image Generation; Aerospace Information
Read the original text (copy or click the link below):
https://cjlcd.lightpublishing.cn/thesisDetails#10.37188/CJLCD.2023-0056
Online Class Incremental Learning for Multi-Pose Point Cloud Targets
Authors:Zhang Runjiang, Guo Jielong, Yu Hui, Lan Hai, Wang Xihao, Wei Xian
Abstract:In light of the current phenomenon where incremental learning targets are all fixed poses, this paper considers a stricter setting, namely online class incremental learning for multi-pose targets, and proposes a pose-agnostic replay method to mitigate catastrophic forgetting when facing multi-pose targets in online class incremental learning. First, 2D/3D targets are point-cloud processed to facilitate the extraction of effective geometric information; second, the network is improved for translation and rotation invariance based on SE (d)(d=2,3), allowing it to extract richer geometric information and reducing the model’s susceptibility to target pose during each task; finally, specific samples are sampled for replay based on loss changes to alleviate catastrophic forgetting. Experimental results show that when facing fixed pose targets like MNIST and CIFAR-10, the final average accuracy of this method reaches 88% and 42.6%, respectively, which is comparable to the results of contrast methods, but the final average forgetting rate is significantly better than that of contrast methods, reduced by about 3% and 15%, respectively. When facing multi-pose targets like RotMNIST and trCIFAR-10, this method still maintains good performance in fixed pose targets, largely unaffected by target poses. Furthermore, its performance on the 3D dataset ModelNet40 remains stable. The proposed method can mitigate catastrophic forgetting without being affected by target poses in online class incremental learning, demonstrating excellent stability and plasticity.
Keywords:Online Class Incremental Learning; Catastrophic Forgetting; Pose-Agnostic Replay; Invariance; Point Cloud Classification
Read the original text (copy or click the link below):
https://cjlcd.lightpublishing.cn/thesisDetails#10.37188/CJLCD.2022-0419
Image Enhancement
Underwater Image Enhancement Method Combining Feature Fusion and Physical Correction
Authors: Wang Dexing, Yang Yurui, Yuan Hongchun, Gao Kai
Abstract: To address the severe color cast and low contrast quality issues in underwater images caused by light absorption and scattering phenomena, this paper proposes a lightweight feature fusion network combined with a multi-color model correction underwater image enhancement method. First, a feature fusion network using self-constructed blocks instead of convolution layers is employed to correct color casts in underwater images, where the improved feature fusion module reduces the destruction of spatial structure by fully connected layers, preserves spatial features, and reduces the parameter count of the module. Simultaneously, an improved attention module performs parallel pooling calculations to extract texture details of feature maps while preserving background information. Then, a multi-color model correction module is used to correct based on pixel relationships, further reducing color casts and improving contrast and brightness. Experimental results show that compared to the latest image enhancement methods, this method achieves an average improvement of 9.3%, 3.7%, and 2.3% in NRMSE, PSNR, and SSIM evaluation metrics, respectively, on reference image datasets. On no-reference image datasets, this method’s average values for UCIQE, IE, and NIQE are improved by 6.0%, 2.9%, and 4.5%, respectively. Comprehensive subjective perception and objective evaluation indicate that this method can correct underwater image color casts, enhance contrast and brightness, and improve image quality.
Keywords: Image Processing; Neural Networks; Attention Mechanism; Color Models; Encoder-Decoder Structure
Read the original text (copy or click the link below):
https://cjlcd.lightpublishing.cn/thesisDetails#10.37188/CJLCD.2022-0382
Neural Network Non-Uniformity Correction Algorithm Improved by Side Window Filtering
Authors:Li Mingqing, Wang Yuqing, Sun Haijiang
Abstract:In the non-uniformity correction problem of infrared focal plane array detectors (IRFPA), traditional neural network algorithms may result in blurred image edges, low contrast, and ghosting phenomena. To address these issues, this paper proposes a neural network non-uniformity correction algorithm improved by side window filtering. This algorithm first applies side window filtering to the input image to obtain the desired image, effectively removing non-uniformity noise while preserving target edge details, thereby enhancing image quality. On this basis, a saturation non-linear function is used to suppress local divergence of correction parameters, effectively preventing ghosting issues in corrected images. Experimental results show that the proposed algorithm can effectively remove non-uniformity noise from images without obvious ghosting phenomena. In three sets of test image sequences, the average image roughness decreases by 30.17%. The maximum time taken to continuously process 400 frames of image sequences on the experimental computer is 37.4170 s, which is 95.05% less than the time taken by the contrast algorithm based on bilateral filtering and 45.81% less than the time taken by the contrast algorithm based on wavelet principal component analysis. This algorithm demonstrates significant advantages in both non-uniformity correction effects and algorithm running efficiency, providing new research ideas for achieving real-time non-uniformity correction on low-power mobile platforms.
Keywords:Infrared Focal Plane Array; Non-Uniformity Correction; Side Window Filtering; Neural Networks
Read the original text (copy or click the link below):
https://cjlcd.lightpublishing.cn/thesisDetails#10.37188/CJLCD.2022-0423
Image Mixed Distortion Correction Method Based on Improved U-Net Network
Authors:Song Wei, Shi Libiao, Geng Lijia, Ma Zhenling, Du Yanling
Abstract:Image geometric distortion correction is a key preprocessing step for many computer vision applications. Currently, deep learning-based geometric distortion correction methods mainly address single distortion correction problems. To this end, this paper proposes an improved U-Net network for image mixed distortion correction. First, a method for constructing a mixed distortion image dataset is proposed to address the scarcity of training datasets and the single distortion type issue. Second, a U-Net network combined with a spatial attention mechanism is employed for image feature extraction and distortion coordinate map reconstruction, transforming the image correction problem into a prediction problem of pixel-by-pixel coordinate displacement changes, and a loss function combining coordinate difference loss and image resampling loss is designed to effectively improve correction accuracy. Finally, ablation experiments verify the performance of each module in this method. Compared to the latest deep learning-based distortion correction methods, experimental results show that this method performs well in both quantitative indicators and subjective evaluations, with an average absolute error of 0.2519 for spatial coordinate correction of distorted images. Calibration experiments on optical images obtained from GoPro cameras further validate the effectiveness of this method for correcting distorted images.
Keywords:Mixed Distortion Correction; U-Net; Spatial Attention; Coordinate Difference Loss; Resampling Loss
Read the original text (copy or click the link below):
https://cjlcd.lightpublishing.cn/thesisDetails#10.37188/CJLCD.2022-0387
Image Segmentation
FRKDNet: Knowledge Distillation-Based Feature Refinement Semantic Segmentation Network
Authors:Jiang Shiyi, Xu Yang, Li Danyang, Fan Runze
Abstract:Traditional semantic segmentation knowledge distillation methods still face issues of incomplete knowledge distillation and insignificant feature information transfer, and the complexity of knowledge transfer from the teacher network can easily lead to the loss of feature positional information. To address these issues, this paper proposes a knowledge distillation-based feature refinement semantic segmentation model, FRKDNet. First, a feature refinement method is designed to separate foreground content from background noise based on the characteristics of foreground features and background noise, filtering out pseudo-knowledge from the teacher network and transmitting more accurate feature content to the student network, thereby enhancing feature representation capability. Meanwhile, inter-class distance and intra-class distance are extracted in the implicit encoding of feature space to obtain the corresponding feature coordinate mask, enabling the student network to simulate the positional information of features and minimize the gap with the teacher network’s feature positions, and the distillation loss is calculated separately for the student network, thereby improving the segmentation accuracy of the student network and assisting it in converging more quickly. Finally, excellent segmentation performance is achieved on the public datasets Pascal VOC and Cityscapes, with MIoU reaching 74.19% and 76.53%, respectively, improving by 2.04% and 4.48% compared to the original student network. This method demonstrates better segmentation performance and robustness compared to mainstream methods, providing a new approach for semantic segmentation knowledge distillation.
Keywords:Semantic Segmentation; Neural Networks; Knowledge Distillation; Feature Refinement; Deep Learning
Read the original text (copy or click the link below):
https://cjlcd.lightpublishing.cn/thesisDetails#10.37188/CJLCD.2023-0010
Fast Two-Dimensional Cumulative Residual Tsallis Entropy Threshold Segmentation Method
Authors:Huang Cong, Zou Yaobin
Abstract:For images with a bimodal gray histogram, traditional two-dimensional histogram threshold segmentation methods are effective, but they perform poorly when the gray histogram presents a non-peak, unimodal, or multimodal pattern. Considering the advantages of density continuity and morphological uniformity of the two-dimensional survival function obtained through two-dimensional histogram mapping, this paper proposes a fast two-dimensional cumulative residual Tsallis entropy threshold segmentation method based on the two-dimensional survival function of images. This method first constructs a two-dimensional survival function based on the two-dimensional histogram, then defines a two-dimensional cumulative residual Tsallis entropy objective function for calculating the segmentation threshold based on the two-dimensional survival function. The time complexity for calculating the objective function is reduced to O(L²) through a recursive algorithm. Finally, the optimal threshold vector for segmentation is obtained based on the recursive form of the two-dimensional cumulative residual Tsallis entropy criterion. The proposed method is compared with two fast two-dimensional threshold segmentation methods, two clustering segmentation methods, and one active contour segmentation method based on time and misclassification error (ME) metrics on 26 synthetic images and 76 real-world images. Experimental results show that, compared to the second-best performing method, this method reduces the average time by 0.013 s and the average ME value by 0.051–0.089 in both synthetic and real-world images. The proposed fast two-dimensional cumulative residual Tsallis entropy threshold segmentation method not only outperforms the five compared methods in computational efficiency but also shows significant advantages in segmentation adaptability and accuracy.
Keywords:Threshold Segmentation; Two-Dimensional Histogram; Two-Dimensional Survival Function; Cumulative Residual Tsallis Entropy; Fast Recursive Algorithm
Read the original text (copy or click the link below):
https://cjlcd.lightpublishing.cn/thesisDetails#10.37188/CJLCD.2022-0427
Supervised by: Zhang Ying, Zhao Yang
Editor: Zhao Wei
Banner Promotion
Some readers say they often do not receive the articles pushed by the public account.
Finally, I discovered a secret!
It turns out, if you do not “star”⭐️, you won’t see me!
Currently, there is an effective solution: Click the public account name “China Optics” at the top of the article, which will take you to an interface. Click the icon in the upper right corner “···“, a new interface will pop up, and click “Set as Starred” (method shown below).
Welcome research groups to submit promotional articles
Please scan the code to contact the duty editor
Click here 👇 to follow me
Light up “Like” and “Looking” so you won’t miss article updates