In the endless hot rolling process, camber is a significant defect where camber is flatness asymmetries produced by nonuniform distribution of rolling pressure. This kind of defect may not only cause clogging of the finishing mill, but also visible defects, such as bar edge folds, edge cracks, holes, and scrapes. It is important to measure the exact camber values of the overall shape and head/tail part of steel bars to reduce camber for hot-rolled steel bars. We introduce a vision-based camber measurement system based on two-area-scan charge-coupled device (CCD) cameras and proposes a camber-detection algorithm to achieve the optimal cutting-line, the overall bar shape, and camber values from the obtained image. The proposed algorithm consists of three parts: optimal cutting-line detection, image-stitching, and camber detection. The optimal cutting-line detection part determines the cutting line of the head/tail part of the bar for continuous joining between steel bars. The image-stitching part obtains the overall shape of a hot-rolled steel bar with continuous image sequence obtained by two-area-scan CCD cameras. The camber-detection part measures head camber, tail camber, and whole camber of the steel bar. Through the proposed camber-detection algorithm, the system not only satisfies a time constraint for detecting the camber of the steel bar but also obtains the optimal cutting line, the overall bar shape, and the camber values.

Vision-based camber measurement system in the endless hot rolling process

In order to realize autonomous landing of the unmanned aerial vehicle (UAV) in power patrolling, a visual method vision based on Faster Regions with Convolutional Neural Network (Faster R-CNN) for UAVs is studied. In this paper, we design the landing sign of the combination of concentric circles and pentagon, and propose the Faster R-CNN recognition algorithm which can be used to identify the target sign. Faster R-CNN successfully identifying the landing mark is the most important step for the UAV autonomous landing. Then, the estimation algorithm of position and direction based on vision is proposed. Position and direction for the UAV landing can be obtained based on least squares ellipse fitting and Shi-Tomasi corner detection method after the landing sign is effectively identified by Faster R-CNN. The experimental results show that it can achieve recognition speed of nearly 81 millisecond each frame and 97.8% accuracy by using Faster R-CNN for detection and identification. The proposed method has better identification accuracy compared with three target identification methods, the Support Vector Machine (SVM) classification, the Back Propagation (BP) neural network and You Only Look Once (YOLO) based on deep learning. The position and direction estimation error of the vision algorithm is within the allowable range, and it can meet the UAV real-time landing requirements.

Identification of autonomous landing sign for unmanned aerial vehicle based on faster regions with convolutional neural network

Character recognition is an art of detecting and recognizing both alphabets and numbers from input images thereby making tremendous advancements in automation process. There are many methods available for character recognition namely OCR by matrix matching, fuzzy logic etc. In this paper, concepts of machine learning are used to develop an efficient algorithm for character recognition. Different feature extraction techniques are designed so as to train the character recognition network. The proposed classifier yields high accuracy than the existing techniques and can be used in high accuracy end to end system.

A SVM based character recognition system

In visual object recognition problems essential to surveillance and navigation problems in a variety of military and civilian use cases, low-resolution and low-quality images present great challenges to this problem. Recent advancements in deep learning based methods like EDSR/VDSR have boosted pixel domain image super-resolution (SR) performances significantly in terms of signal to noise ratio(SNR)/ mean square error(MSE) metrics of the super-resolved image. However, these pixel domain signal quality metrics may not directly correlate to the machine vision tasks like key points detection and object recognition. In this work, we develop a machine vision tasks-friendly super-resolution technique which enhances the gradient images and associated features from the low-resolution images that benefit the high level machine vision tasks. Here, a residual learning deep neural network based gradient image super-resolution solution is developed with scale space adaptive network depth, and simulation results demonstrate the performance gains in both gradient image quality as well as key points repeatability.

Gradient Image Super-resolution for Low-resolution Image Recognition

In color aerial images, shadows appearing in images may cause false color tones, loss of feature information, and shape distortion of objects, which affects further image processing tasks. Hence, shadow detection is one of the challenging tasks in image processing. In this paper, we will study basics of shadows and different types of shadows which can appear in the images. Significant literature review is carried out on existing shadow detection techniques. Simple and efficient chromaticity-based approach for shadow detection is applied on color aerial images. The technique is applied in YIQ color space that decouple luminance and chromaticity. The technique is further extended by us to further enhance the shadow detection results using proposed post processing methods. The visual results shows effectiveness of our approach in comparison with the results obtained using the original approach without post processing.

Shadow detection based on RGB and YIQ color models in color aerial images

Corner points are formed from two or more edges and edges usually define the boundary between two different objects or parts of the same objects. In this novel we discuss the theory of the Harris corner detection and indicate its disadvantage. Then it proposes an improved algorithm of Harris detection algorithm based on the neighboring point eliminating method. It reduces the time of the detection, and makes the corners distributing more homogenous so that avoids too many corners stay together. Experimental results show that the algorithm can detect corner more equality distributing, and can be used in some fact applications such as image registration well. Key words-Corner point; Harris corner; neighboring

An Improved Algorithm for Harris Corner Detection

Corner points are interesting as they are formed from two or more edges and edges usually define the boundary between two different objects or parts of the same objects. In this paper we discuss the theory of the Harris corner detection and point out its disadvantage. Then it proposes an adaptive algorithm for Harris corner detection based on the neighboring point eliminating method. It reduces the amount of the calculation, and makes the corners distributing well-proportion so that avoids too many corners stay together. Experimental results show that the algorithm can detect corners more uniform distribution, and can be used in some fact applications such as image registration well. Key words-Harris corner; neighboring point;

An Adaptive Algorithm for Harris Corner Detection

Convolutional network models have been widely used in image segmentation. However, there are many types of boundary contour features in medical images which seriously affect the stability and accuracy of image segmentation models, such as the ambiguity of tumors, the variability of lesions, and the weak boundaries of fine blood vessels. In this paper, in order to solve these problems we first introduce the dual-tree complex wavelet scattering transform module, and then innovatively propose a learning scattering wavelet network model. In addition, a new improved active contour loss function is further constructed to deal with complex segmentation. Finally, the equilibrium coefficient of our model is discussed. Experiments on the BraTS2020 dataset show that the LSW-Net model has improved the Dice coefficient, accuracy, and sensitivity of the classic FCN, SegNet, and At-Unet models by at least 3.51%, 2.11%, and 0.46%, respectively. In addition, the LSW-Net model still has an advantage in the average measure of Dice coefficients compared with some advanced segmentation models. Experiments on the DRIVE dataset prove that our model outperforms the other 14 algorithms in both Dice coefficient and specificity measures. In particular, the sensitivity of our model provides a 3.39% improvement when compared with the Unet model, and the model’s effect is obvious.

/pdf/lsw-net-a-learning-scattering-wavelet-network-for-brain-76ioehc5.pdf

LSW-Net: A Learning Scattering Wavelet Network for Brain Tumor and Retinal Image Segmentation

Ischemic strokes are the most common acute brain disorder, and seriously threaten patients’ lives. In order to help physicians determine the location of ischemic stroke lesions and other information as early as possible, many scholars have used convolutional neural networks and Transformer segmentation networks to segment lesions on CT perfusion images. However, convolutional neural networks are not capable of extracting spatial information sufficiently, which leads to loss of effective lesion information. In addition, the global attention mechanism module of Transformer is computationally intensive at runtime, which is not suitable for use in high-resolution input and intensive prediction tasks. We designed a DSE-ResNet module to solve these problems to establish spatial channel information correlation. Then we innovatively propose the Pool-UNet model, which combines the Poolformer structure with a convolutional neural network. It can efficiently model the global context and learn multi-scale features while maintaining a grasp of the lowlevel details. The segmentation results on the ISLES-2018 dataset show that PoolUNet achieves 67.82% precision, 56.54% recall, 56.04% Dice coefficient, and 21.14 mm Haushofer distance. Compared with the classical UNet, R2UNet, and TransUNet 3 segmentation models, Pool-UNet improved at least 0.26%, 1.52%, 1.07%, and 0. 17mm in accuracy, recall, Dice coefficient, and Hausdorff distance, respectively. Pool-UNet has a competitive advantage over other classical and advanced medical segmentation algorithms.

Zhiyong Ye

Papers

An Improved Algorithm for Harris Corner Detection

An Adaptive Algorithm for Harris Corner Detection

LSW-Net: A Learning Scattering Wavelet Network for Brain Tumor and Retinal Image Segmentation

Pool-UNet: Ischemic Stroke Segmentation from CT Perfusion Scans Using Poolformer UNet