Area-based non-maximum suppression algorithm for multi-object fault detection

Unmanned aerial vehicle(UAV)photography has become the main power system inspection method;however,automated fault detection remains a major challenge.Conventional algorithms encounter difficulty in processing all the detected objects in the power transmission lines simultaneously.The object detection method involving deep learning provides a new method for fault detection.However,the traditional non-maximum suppression(NMS)algorithm fails to delete redundant annotations when dealing with objects having two labels such as insulators and dampers.In this study,we propose an area-based non-maximum suppression(A-NMS)algorithm to solve the problem of one object having multiple labels.The A-NMS algorithm is used in the fusion stage of cropping detection to detect small objects.Experiments prove that A-NMS and cropping detection achieve a mean average precision and recall of 88.58%and 91.23%,respectively,in case of the aerial image datasets and realize multi-object fault detection in aerial images.

一种改进的基于偏微分方程图像的降噪算法

针对图像去噪过程中存在边缘保持与噪声抑制之间的矛盾,提出了一种改进的基于偏微分模型的图像去噪算法.引入了正则化、绝对差值排序检测法,结合Chao和Tsai模型,构造了一种新的扩散系数函数,兼具了正则化解决方程病态问题和绝对差值排序检测法有效区分噪声与边缘的优点.实验结果表明:与其他基于常用偏微分模型的去噪算法相比,所提算法能更加有效地去除噪声,保留了更多的细节信息,提高了图像去噪的信噪比.

一种多层特征融合的人脸检测方法

由于姿态、光照、尺度等原因,卷积神经网络需要学习出具有强判别力的特征才能应对复杂场景下的人脸检测问题。受卷积神经网络中特定特征层感受野大小限制,单独一层的特征无法应对多姿态多尺度的人脸,为此提出了串联不同大小感受野的多层特征融合方法用于检测多元化的人脸;同时,通过引入加权降低得分的方法,改进了目前常用的非极大值抑制算法,用于处理由于遮挡造成的相邻人脸的漏检问题。在FDDB和WiderFace两个数据集上的实验结果显示,文中提出的多层特征融合方法能显著提升检测结果,改进后的非极大值抑制算法能够提升相邻人脸之间的检测准确率。

基于深度学习的自顶向下人体姿态估计算法

针对自顶向下的人体姿态估计算法出现的目标框定位错误问题和冗余检测问题,提出了一种基于深度学习的自顶向下人体姿态估计算法。设计了对称空间变换网络与单人姿态估计网络相连接,以从不准确的人体边界框中提出高质量的人体目标框,并且引入了参数化姿态非极大值抑制消除了冗余的姿态估计,应用消除规则对相似的姿态进行消除,得到唯一的人体姿态估计结果。在公共人体姿态估计数据集MPII上选取部分数据集进行训练和测试,实验结果表明所提出的方法能够准确地检测出人体关键点,有效地提高了人体姿态估计的准确率,且能够适应人员密集、存在遮挡的复杂场景。

Weighted total variation using split Bregman fast quantitative susceptibility mapping reconstruction method

An ill-posed inverse problem in quantitative susceptibility mapping （QSM） is usually solved using a regularization and optimization solver, which is time consuming considering the three-dimensional volume data. However, in clinical diagnosis, it is necessary to reconstruct a susceptibility map efficiently with an appropriate method. Here, a modified QSM reconstruction method called weighted total variation using split Bregman （WTVSB） is proposed. It reconstructs the susceptibility map with fast computational speed and effective artifact suppression by incorporating noise-suppressed data weighting with split Bregman iteration. The noise-suppressed data weighting is determined using the Laplacian of the calculated local field, which can prevent the noise and errors in field maps from spreading into the susceptibility inversion. The split Bregman iteration accelerates the solution of the Ll-regularized reconstruction model by utilizing a preconditioned conjugate gradient solver. In an experiment, the proposed reconstruction method is compared with truncated k-space division （TKD）, morphology enabled dipole inversion （MEDI）, total variation using the split Bregman （TVSB） method for numerical simulation, phantom and in vivo human brain data evaluated by root mean square error and mean structure similarity. Experimental results demonstrate that our proposed method can achieve better balance between accuracy and efficiency of QSM reconstruction than conventional methods, and thus facilitating clinical applications of QSM.

Accurate Registration of Remote Sensing Images Based on Local Optimal Transformation

As the basic work of image stitching and object recognition,image registration played an important part in the image processing field.Much previous work in registration accuracy and realtime performance progressed very slowly,especially in registrating images with line feature.An innovative method for image registration based on lines is proposed,it can effectively improve the accuracy and real-time performance of image registration.The line feature can deal with some registration problems where point feature does not work.Our registration process is divided into two parts.The first part determines the rough registration transformation relation between reference image and test image.Then the similarity degree among different transformation and modified nonmaximum suppression(MNMS)algorithms are obtained,which produce local optimal solution to optimize the rough registration transformation.The final optimal registration relation can be obtained from two registration parts according to the match scores.The experimental results show that the proposed method makes a more accurate registration relation and performs better in real-time situation.

YOLO-Banana:An Effective Grading Method for Banana Appearance Quality

The increasing trend towards independent fruit packaging demands a high appearance quality of individually packed fruits.In this paper,we propose an improved YOLOv5-based model,YOLO-Banana,to effectively grade banana appearance quality based on the number of banana defect points.Due to the minor and dense defects on the surface of bananas,existing detection algorithms have poor detection results and high missing rates.To address this,we propose a densitybased spatial clustering of applications with noise(DBSCAN)and K-means fusion clustering method that utilizes refined anchor points to obtain better initial anchor values,thereby enhancing the network’s recognition accuracy.Moreover,the optimized progressive aggregated network(PANet)enables better multi-level feature fusion.Additionally,the non-maximum suppression function is replaced with a weighted non-maximum suppression(weighted NMS)function based on distance intersection over union(DIoU).Experimental results show that the model’s accuracy is improved by 2.3%compared to the original YOLOv5 network model,thereby effectively grading the banana appearance quality.

Research on Pedestrian Detection Technology Based on MSR and Faster R-CNN

In order to avoid the problem of poor illumination characteristics and inaccurate positioning accuracy, this paper proposed a pedestrian detection algorithm suitable for low-light environments. The algorithm first applied the multi-scale Retinex image enhancement algorithm to the sample pre-processing of deep learning to improve the image resolution. Then the paper used the faster regional convolutional neural network to train the pedestrian detection model, extracted the pedestrian characteristics, and obtained the bounding boxes through classification and position regression. Finally, the pedestrian detection process was carried out by introducing the Soft-NMS algorithm, and the redundant bounding box was eliminated to obtain the best pedestrian detection position. The experimental results showed that the proposed detection algorithm achieves an average accuracy of 89.74% on the low-light dataset, and the pedestrian detection effect was more significant.

Restricted Hysteresis Reduce Redundancy in Edge Detection

In edge detection algorithms, there is a common redundancy problem, especially when the gradient direction is close to -135°, -45°, 45°, and 135°. Double edge effect appears on the edges around these directions. This is caused by the discrete calculation of non-maximum suppression. Many algorithms use edge points as feature for further task such as line extraction, curve detection, matching and recognition. Redundancy is a very important factor of algorithm speed and accuracy. We find that most edge detection algorithms have redundancy of 50% in the worst case and 0% in the best case depending on the edge direction distribution. The common redundancy rate on natural images is approximately between 15% and 20%. Based on Canny’s framework, we propose a restriction in the hysteresis step. Our experiment shows that proposed restricted hysteresis reduce the redundancy successfully.

CARVING-DETC: A network scaling and NMS ensemble for Balinese carving motif detection method

Balinese carvings are cultural objects that adorn sacred buildings. The carvings consist of several motifs,each representing the values adopted by the Balinese people. Detection of Balinese carving motifs ischallenging due to the unavailability of a Balinese carving dataset for detection tasks, high variance,and tiny-size carving motifs. This research aims to improve carving motif detection performance onchallenging Balinese carving motifs detection task through a modification of YOLOv5 to support adigital carving conservation system. We proposed CARVING-DETC, a deep learning-based Balinesecarving detection method consisting of three steps. First, the data generation step performs dataaugmentation and annotation on Balinese carving images. Second, we proposed a network scalingstrategy on the YOLOv5 model and performed non-maximum suppression (NMS) on the modelensemble to generate the most optimal predictions. The ensemble model utilizes NMS to producehigher performance by optimizing the detection results based on the highest confidence score andsuppressing other overlap predictions with a lower confidence score. Third, performance evaluation onscaled-YOLOv5 versions and NMS ensemble models. The research findings are beneficial in conservingthe cultural heritage and as a reference for other researchers. In addition, this study proposed a novelBalinese carving dataset through data collection, augmentation, and annotation. To our knowledge,it is the first Balinese carving dataset for the object detection task. Based on experimental results,CARVING-DETC achieved a detection performance of 98%, which outperforms the baseline model.