Locating Objects Without Bounding Boxes

Forward train detection is of great significance to improve train safety. In this paper, a vision-based method is proposed for forward train detection. The proposed method based on CenterNet, a novel object detection network, to realize accurate and fast forward train detection. The forward train detection network is divided into two stages: downsampling stage and center points generation stage. In downsampling stage, a full convolution network is performed to downsample the image, meanwhile to extract the feature of the image. In the center points generation stage, three branch networks are used to predict the bounding box of the forward train, including heatmaps generation branch, center offset regression branch and forward train bounding box size regression branch. Experiments results show that the proposed method can detect the forward train well and achieve 30.7% Average Precision (AP) with a 47.6 Frames Per Second (FPS).

An object detector based on convolutional neural network (CNN) has been widely used in the field of computer vision because of its simplicity and efficiency. The average accuracy of CNN model detection results in the object detector is greatly affected by the loss function. The precision of the localization algorithm in the loss function is the main factor affecting the result. Based on the complete intersection over union (CIoU) loss function, an improved penalty function is proposed to improve the localization accuracy. Specifically, the algorithm more comprehensively considers matching bounding boxes between prediction with ground truth, using the proportional relationship of the aspect ratio from both bounding boxes. Under the same aspect ratio of the two bounding boxes, the influence factors of the prediction box on localization accuracy were considered. In this way, the function of the penalty function is strengthened, and localization accuracy of the network model improved. This loss function is called Improved CIoU (ICIoU). Experiments on the Udacity, PASCAL VOC, and MS COCO datasets have demonstrated the effectiveness of ICIoU in improving localization accuracy of network models by using the one-stage object detector YOLOv4. Compared with CIoU, the proposed ICIoU improved average precision (AP) by 0.57% and AP75 by 0.12% on Udacity, AP by 0.26% and AP75 by 1.28% on PASCAL VOC, and AP by 0.06% and AP75 by 0.65% on MS COCO.

In this paper, a novel method to deal with the semantic segmentation of very high resolution remote sensing data is presented. Recent advances in deep learning (DL), especially convolutional neural networks (CNNs) and fully convolutional networks (FCNs), have shown outstanding performances in this task. However, the map accuracy depends on the quantity and quality of ground truth (GT) used to train them. At the same time, probabilistic graphical models (PGMs) have sparked even more interest in the past few years, because of the ever-growing need for structured predictions. The novel method proposed in this paper combines DL and PGMs to perform remote sensing image classification. FCNs can be exploited to deal with multiscale data through the integration with a hierarchical Markov model. The marginal posteri- or mode (MPM) criterion for inference is used in the proposed framework. Experimental validation is conducted on the ISPRS 2D Semantic Labeling Challenge Vaihingen dataset. The results are significant, as the proposed method has a higher recall than the standard FCNs considered and allows mitigating the impact of incomplete or suboptimal GT, especially with regard to the discrimination of minoritary classes.

In this paper we propose the Localized Quality Aware Image Denoising (LQAID) technique for image denoising using deep convolutional neural networks (CNNs). LQAID relies on local quality estimates over global cues like noise standard deviation since the perceptual quality of a noisy image is typically spatially varying. Specifically, we use localized quality maps generated using DistNet, a spatial quality map estimation method. These quality maps are used to augment the noisy image and guide the denoising process. The augmented noisy image is denoised using a deep fully convolutional network (FCN) trained using mean square error (MSE) as the loss function. The proposed approach shows state-of-the-art performance both qualitatively and quantitatively on two vision datasets: TID 2008 and BSD500. We also show that the proposed approach possesses excellent generalization ability. Lastly, the proposed approach is completely blind since it neither requires information about the strength of the additive noise nor does it try to explicitly estimate it.

Optical pressure sensors based plantar image segmenting using an improved fully convolutional network

In order to improve the performance of the target detection model based on convolutional neural network, and make the bounding box representing the object location contain as complete original target information as possible, this paper improved the loss function of the bounding box regression part of the YOLOv3 network. A new metric, Asymmetric Intersection over Union (AIoU), is proposed to measure the similarity between the prediction box and the truth. Theoretical analysis shows that AIoU based loss function can better guide the direction of bounding box regression after introducing asymmetry. Experiments show that the improved YOLOv3 network achieves higher accuracy and maintains fast convergence in human target detection. The examples of bounding box regression on the test dataset show that the target information loss can be reduced by the proposed method. In addition, distinguishing between the prediction and the truth in calculation may have reference significance for the loss function design of network models for other tasks

To facilitate the management of detonators, an automatic spotting system is proposed for detonator coded characters based on convolutional neural networks. The system contains a multi-scale detection network, a multi-label recognition network and a post-processing layer. An improved fully convolution network (FCN) is designed as the multi-scale detection network to obtain the accurate response map of the detonator image. Two subnetworks are parallel integrated into the FCN to perform a coarse-to-fine detection. An improved Jaccard loss function with a regularization term is defined to train the FCN. The region of interest (ROI) of the detonator image is achieved when the response map is post-processed by the post-processing layer. Finally, a modified multi-label network is used to recognize the detonator coded characters in the ROI. The experimental results indicate that the proposed system achieves a better spotting performance for detonator coded characters than the state-of-the-art text spotting methods in terms of accuracy and efficiency.

Early detection and treatment are crucial for recovery from cancer. Due to rapid advancement in medical imaging techniques, today doctors can scan various body parts from different perspectives to diagnose cancer or see the progress during a course of treatment. Such techniques are less painful and require less preparation and recovery. The recent evolution of AI, particularly deep learning algorithms have shown potential in computer-aided diagnosis of cancer. These include analysis of cancer images for detection of tumor region, prognosis, skin tissue inspection, etc. Convolutional Neural Network (CNN) based image processing techniques are showing promising results in the diagnosis of breast cancer, skin cancer, prostate cancer, lung cancer, etc. Literature shows many other CNN-based models such as feed-forward neural network, Deep CNN, Multi-Crop CNN, fully convolutional networks (FCNs), Deep Fully Convolutional Network (DFCNet), Recurrent neural networks, etc. The CNN focus solely on the image-specific features and hence require a lesser number of input parameters. A CNN model reduces the size of the input image vector without losing the features critical for making an accurate prediction. This makes it easier to process and reduces computational complexity. CNNs are also not sensitive to the object positioning in an image since the approximate position of image motifs relative to others is known. CNN deals with the variations in images through augmentation and other pre-processing techniques. This makes them robust to various straining conditions in images. This chapter presents an overview of a generic CNN architecture followed by a discussion on major advances in CNN models and their applications for cancer image analysis.

We propose a novel, multi-task, fully convolutional network (FCN) architecture for automatic segmentation of brain tumor. This network extracts multi-level contextual information by concatenating hierarchical feature representations extracted from multimodal MR images along with their symmetric-difference images. It achieves improved segmentation performance by incorporating boundary information directly into the loss function. The proposed method was evaluated on the BRATS13 and BRATS15 datasets and compared with competing methods on the BRATS13 testing set. Segmented tumor boundaries obtained were better than those obtained by single-task FCN and by FCN with CRF. The method is among the most accurate available and has relatively low computational cost at test time.

This study aims to enhance the accuracy and efficiency of MRI-based brain tumor diagnosis by leveraging deep learning (DL) techniques applied to multichannel MRI inputs. MRI data were collected from 203 subjects, including 100 normal cases and 103 cases with 13 distinct brain tumor types. Non-contrast T1-weighted (T1w) and T2-weighted (T2w) images were combined with their average to form RGB three-channel inputs, enriching the representation for model training. Several convolutional neural network (CNN) architectures were evaluated for tumor classification, while fully convolutional networks (FCNs) were employed for tumor segmentation. Standard preprocessing, normalization, and training procedures were rigorously followed. The RGB fusion of T1w, T2w, and their average significantly enhanced model performance. The classification task achieved a top accuracy of 98.3% using the Darknet53 model, and segmentation attained a mean Dice score of 0.937 with ResNet50. These results demonstrate the effectiveness of multichannel input fusion and model selection in improving brain tumor analysis. While not yet integrated into clinical workflows, this approach holds promise for future development of DL-assisted decision-support tools in radiological practice.

Land cover is an important variable of the terrestrial ecosystem that provides information for natural resources management, urban sprawl detection, and environment research. To classify land cover with high-spatial-resolution multispectral remote sensing imagery is a difficult problem due to heterogeneous spectral values of the same object on the ground. Fully convolutional networks (FCNs) are a state-of-the-art method that has been increasingly used in image segmentation and classification. However, a systematic quantitative comparison of FCNs on high-spatial-multispectral remote imagery was not yet performed. In this paper, we adopted the three FCNs (FCN-8s, Segnet, and Unet) for Gaofen-2 (GF2) satellite imagery classification. Two scenes of GF2 with a total of 3329 polygon samples were used in the study area and a systematic quantitative comparison of FCNs was conducted with red, green, blue (RGB) and RGB+near infrared (NIR) inputs for GF2 satellite imagery. The results showed that: (1) The FCN methods perform well in land cover classification with GF2 imagery, and yet, different FCNs architectures exhibited different results in mapping accuracy. The FCN-8s model performed best among the Segnet and Unet architectures due to the multiscale feature channels in the upsampling stage. Averaged across the models, the overall accuracy (OA) and Kappa coefficient (Kappa) were 5% and 0.06 higher, respectively, in FCN-8s when compared with the other two models. (2) High-spatial-resolution remote sensing imagery with RGB+NIR bands performed better than RGB input at mapping land cover, and yet the advantage was limited; the OA and Kappa only increased an average of 0.4% and 0.01 in the RGB+NIR bands. (3) The GF2 imagery provided an encouraging result in estimating land cover based on the FCN-8s method, which can be exploited for large-scale land cover mapping in the future.

Farmland is the cornerstone of agriculture and is important for food security and social production. Farmland assessment is essential but traditional methods are usually expensive and slow. Deep learning methods have been developed and widely applied recently in image recognition, semantic understanding, and many other application domains. In this research, we used fully convolutional networks (FCN) as the deep learning model to evaluate farmland grades. Normalized difference vegetation index (NDVI) derived from Landsat images was used as the input data, and the China National Cultivated Land Grade Database within Jiangsu Province was used to train the model on cloud computing. We also applied an image segmentation method to improve the original results from the FCN and compared the results with classical machine learning (ML) methods. Our research found that the FCN can predict farmland grades with an overall F1 score (the harmonic mean of precision and recall) of 0.719 and F1 score of 0.909, 0.590, 0.740, 0.642, and 0.023 for non-farmland, level I, II, III, and IV farmland, respectively. Combining the FCN and image segmentation method can further improve prediction accuracy with results of fewer noise pixels and more realistic edges. Compared with conventional ML, at least in farmland evaluation, FCN provides better results with higher precision, recall, and F1 score. Our research indicates that by using remote sensing NDVI data, the deep learning method can provide acceptable farmland assessment without fieldwork and can be used as a novel supplement to traditional methods. The method used in this research will save a lot of time and cost compared with traditional means.

Steel box girder is the main component of modern long-span cable system bridges. And crack is the primary sign of problems with girders, thus its identification is very important. Lots of works have been done to solve this issue by cutting raw images into small blocks, then segmentation of cracks is treated as classification of blocks. How- ever, this type of method cannot be accurately segment cracks at the pixel level. Based on this deficiency, this paper proposes a new approach for pixel-wise crack identification of steel box girder by fully convolutional network (FCN). FCN is a new kind of network for pixel-wise segmentation, consisting of only convolutional layers, without fully connected layers. This study presents a new FCN, built based on modified U-Net to adapt to crack identification problem. The number of convolutional layers is adjusted to extract more scaled features. Different feature fusion method is proposed to take full advantages of multi-scale convolutional features. Then, a dataset called APESS2018 Steel Girder Crack ID Dataset is employed for training and validating the network. The dataset includes 50 raw images and corresponding pixel-wise label images with a size of 4928 × 3264 pixels, including four classes: cracks, ruler, handwriting and girder surface (i.e., background). Due to the limitation of computation power, the input size of the network is changed as 544 × 544 pixels. Thus, the Crack ID Dataset is necessary to be cropped and regrouped to 2700 sub-images. 94% of these sub-images are train-validating set with 10% validating rate, 6% are testing set. 180 epochs are trained with multi-step learning rate from 0.01 to 0.0001. Finally, the accuracy of proposed method is validated, 98.906% for training accuracy and 97.525% for testing. Different loss functions (Categorical Cross-Entropy loss and Dice Coefficient loss) are discussed to illustrate their influence on results.

There has been a significant increase from 2010 to 2016 in the number of people suffering from spine problems. The automatic image segmentation of the spine obtained from a computed tomography (CT) image is important for diagnosing spine conditions and for performing surgery with computer-assisted surgery systems. The spine has a complex anatomy that consists of 33 vertebrae, 23 intervertebral disks, the spinal cord, and connecting ribs. As a result, the spinal surgeon is faced with the challenge of needing a robust algorithm to segment and create a model of the spine. In this study, we developed a fully automatic segmentation method to segment the spine from CT images, and we compared our segmentation results with reference segmentations obtained by well-known methods. We use a hybrid method. This method combines the convolutional neural network (CNN) and fully convolutional network (FCN), and utilizes class redundancy as a soft constraint to greatly improve the segmentation results. The proposed method was found to significantly enhance the accuracy of the segmentation results and the system processing time. Our comparison was based on 12 measurements: the Dice coefficient (94%), Jaccard index (93%), volumetric similarity (96%), sensitivity (97%), specificity (99%), precision (over segmentation 8.3 and under segmentation 2.6), accuracy (99%), Matthews correlation coefficient (0.93), mean surface distance (0.16 mm), Hausdorff distance (7.4 mm), and global consistency error (0.02). We experimented with CT images from 32 patients, and the experimental results demonstrated the efficiency of the proposed method. Highlights A method to enhance the accuracy of spine segmentation from CT data was proposed. The proposed method uses Convolutional Neural Network via redundant generation of class labels. Experiments show the segmentation accuracy has been enhanced.

As a function of the spatial position of the optical coherence tomography (OCT) image, retinal layer thickness is an important diagnostic indicator for many retinal diseases. Reliable segmentation of the retinal layer is necessary for extracting useful clinical information. However, manual segmentation of these layers is time-consuming and prone to bias. Furthermore, due to speckle noise, low image contrast, retinal detachment, and also irregular morphological features make the automatic segmentation task challenging. To alleviate these challenges, in this paper, we propose a new coarse-fine framework combining the full convolutional network (FCN) with a multiphase level set (named FCN-MLS) for automatic segmentation of nine boundaries in retinal spectral OCT images. In the coarse stage, FCN is used to learn the characteristics of specific retinal layer boundaries and achieve classification of four retinal layers. The boundaries are then extracted and the remaining boundaries are initialized based on a priori information about the thickness of the retinal layer. In order to prevent the overlapping of the segmentation interfaces, a regional restriction technique is used in the multi-phase level to evolve the boundaries to achieve fine nine retinal layers segmentation. Experimental results on 1280 B-scans show that the proposed method can segment nine retinal boundaries accurately. Compared with the manual delineation, the overall mean absolute boundary location difference and the overall mean absolute thickness difference were 5.88 ± 2.38μm and 5.81 ± 2.19μm, which showed a good consistency with manual segmentation by the physicians. Our experimental results also outperform state-of-the-art methods.