Fully Convolutional Network Architecture Research Articles

A Resource-Efficient Embedded Iris Recognition System Using Fully Convolutional Networks

Applications of fully convolutional networks (FCN) in iris segmentation have shown promising advances. For mobile and embedded systems, a significant challenge is that the proposed FCN architectures are extremely computationally demanding. In this article, we propose a resource-efficient, end-to-end iris recognition flow, which consists of FCN-based segmentation and a contour fitting module, followed by Daugman normalization and encoding. To attain accurate and efficient FCN models, we propose a three-step SW/HW co-design methodology consisting of FCN architectural exploration, precision quantization, and hardware acceleration. In our exploration, we propose multiple FCN models, and in comparison to previous works, our best-performing model requires 50× fewer floating-point operations per inference while achieving a new state-of-the-art segmentation accuracy. Next, we select the most efficient set of models and further reduce their computational complexity through weights and activations quantization using an 8-bit dynamic fixed-point format. Each model is then incorporated into an end-to-end flow for true recognition performance evaluation. A few of our end-to-end pipelines outperform the previous state of the art on two datasets evaluated. Finally, we propose a novel dynamic fixed-point accelerator and fully demonstrate the SW/HW co-design realization of our flow on an embedded FPGA platform. In comparison with the embedded CPU, our hardware acceleration achieves up to 8.3× speedup for the overall pipeline while using less than 15% of the available FPGA resources. We also provide comparisons between the FPGA system and an embedded GPU showing different benefits and drawbacks for the two platforms.

A segmentation method combining probability map and boundary based on multiple fully convolutional networks and repetitive training

Cell nuclei image segmentation technology can help researchers observe each cell’s stress response to drug treatment. However, it is still a challenge to accurately segment the adherent cell nuclei. At present, image segmentation based on a fully convolutional network (FCN) is attracting researchers’ attention. We propose a multiple FCN architecture and repetitive training (M-FCN-RT) method to learn features of cell nucleus images. In M-FCN-RT, the multiple FCN (M-FCN) architecture is composed of several single FCNs (S-FCNs) with the same structure, and each FCN is used to learn the specific features of image datasets. In this paper, the M-FCN contains three FCNs; FCN1-2, FCN3 and FCNB. FCN1-2 and FCN3 are respectively used to learn the spatial features of cell nuclei for generating probability maps to indicate nucleus regions of an image; FCNB (boundary FCN) is used to learn the edge features of cell nuclei for generating the nucleus boundary. For the training of each FCN, we propose a repetitive training (RT) method to improve the classification accuracy of the model. To segment cell nuclei, we finally propose an algorithm combining the probability map and boundary (PMB) to segment the adherent nuclei. This paper uses a public opening nucleus image dataset to train, verify and evaluate the proposed M-FCN-RT and PMB methods. Our M-FCN-RT method achieves a high Dice similarity coefficient (DSC) of 92.11%, 95.64% and 87.99% on the three types of sub-datasets respectively for probability maps. In addition, segmentation experimental results show the PMB method is more effective and efficient compared with other methods.

Automatic liver segmentation by integrating fully convolutional networks into active contour models.

Automatic and accurate three-dimensional (3D) segmentation of liver with severe diseases from computed tomography (CT) images is a challenging task. Fully convolutional networks (FCNs) have emerged as powerful tools for automatic semantic segmentation, with multiple potential applications in medical imaging. However, the use of a large receptive field and multiple pooling layers in the network leads to poor localization around object boundaries. The network usually makes pixel-wise prediction independently, making it difficult to respect local label consistency and enforce the smoothness of the object boundary. We have developed an automatic liver segmentation method based on a novel framework that integrates fully convolutional network predictions into active contour models (ACM). We use only a single network architecture to generate a pixel label map containing spatial regional information (foreground and background) as well as layered boundary information. We exploit the structured network outcome to define an external constraint force of active contour models. A unique property of the designed force is that both its strength and direction are adaptive to its position and relative distance to the object boundary. The resulting integrated active contour models have the advantages of incorporating both high-level and low-level image information simultaneously, while enforcing the smoothness of the contour. Because the external constraint force can push the evolving contour to the liver boundary and exists everywhere in the image domain, it allows us to place the initial contour far away from the liver boundary. It potentially allows us to control the evolution of the contour in order to preserve the topology of the liver. We have trained and evaluated our model on 73 liver CT scans from a clinic study. The integrated ACM model yields mean dice coefficients (DICE) 95.8±1.4 (%). Without further fine-tuning the network weights for two independent datasets, it yields mean DICE 96.2±0.9 (%) for the SLIVER07 training dataset, and mean DICE 94.3±2.7 (%) for the LiTS training dataset. In comparison with FCN alone model, the integrated ACM model yields improvements in terms of surface distance and DICE values for almost all the cases. Furthermore, the initialization of the active contour can be very far away from the liver boundary. Experimental results for segmenting livers (with severe diseases on CT images resulting in shape and density abnormalities) have revealed that our proposed model improves segmentation results in comparison with FCN alone. Without further fine-tuning the network weights for two independent datasets, the model is capable of handling image variations from different datasets due to its inherent deformable nature. It is relatively easy to integrate more advanced (either existing or future) FCN architecture into our framework to further improve the segmentation performance.

Fully Convolutional Networks and Geographic Object-Based Image Analysis for the Classification of VHR Imagery

Land cover Classified maps obtained from deep learning methods such as Convolutional neural networks (CNNs) and fully convolutional networks (FCNs) usually have high classification accuracy but with the detailed structures of objects lost or smoothed. In this work, we develop a methodology based on fully convolutional networks (FCN) that is trained in an end-to-end fashion using aerial RGB images only as input. Skip connections are introduced into the FCN architecture to recover high spatial details from the lower convolutional layers. The experiments are conducted on the city of Goma in the Democratic Republic of Congo. We compare the results to a state-of-the art approach based on a semi-automatic Geographic object image-based analysis (GEOBIA) processing chain. State-of-the art classification accuracies are obtained by both methods whereby FCN and the best baseline method have an overall accuracy of 91.3% and 89.5% respectively. The maps have good visual quality and the use of an FCN skip architecture minimizes the rounded edges that is characteristic of FCN maps. Additional experiments are done to refine FCN classified maps using segments obtained from GEOBIA generated at different scale and minimum segment size. High OA of up to 91.5% is achieved accompanied with an improved edge delineation in the FCN maps, and future work will involve explicitly incorporating boundary information from the GEOBIA segmentation into the FCN pipeline in an end-to-end fashion. Finally, we observe that FCN has a lower computational cost than the standard patch-based CNN approach especially at inference.

An Exploration of Deep-Learning Based Phenotypic Analysis to Detect Spike Regions in Field Conditions for UK Bread Wheat.

Wheat is one of the major crops in the world, with a global demand expected to reach 850 million tons by 2050 that is clearly outpacing current supply. The continual pressure to sustain wheat yield due to the world's growing population under fluctuating climate conditions requires breeders to increase yield and yield stability across environments. We are working to integrate deep learning into field-based phenotypic analysis to assist breeders in this endeavour. We have utilised wheat images collected by distributed CropQuant phenotyping workstations deployed for multiyear field experiments of UK bread wheat varieties. Based on these image series, we have developed a deep-learning based analysis pipeline to segment spike regions from complicated backgrounds. As a first step towards robust measurement of key yield traits in the field, we present a promising approach that employ Fully Convolutional Network (FCN) to perform semantic segmentation of images to segment wheat spike regions. We also demonstrate the benefits of transfer learning through the use of parameters obtained from other image datasets. We found that the FCN architecture had achieved a Mean classification Accuracy (MA) >82% on validation data and >76% on test data and Mean Intersection over Union value (MIoU) >73% on validation data and and >64% on test datasets. Through this phenomics research, we trust our attempt is likely to form a sound foundation for extracting key yield-related traits such as spikes per unit area and spikelet number per spike, which can be used to assist yield-focused wheat breeding objectives in near future.

Deep gated attention networks for large-scale street-level scene segmentation

Street-level scene segmentation aims to label each pixel of street-view images into specific semantic categories. It has been attracting growing interest due to various real-world applications, especially in the area of autonomous driving. However, this pixel-wise labeling task is very challenging under the complex street-level scenes and large-scale object categories. Motivated by the scene layout of street-view images, in this work we propose a novel Spatial Gated Attention (SGA) module, which automatically highlights the attentive regions for pixel-wise labeling, resulting in effective street-level scene segmentation. The proposed module takes as input the multi-scale feature maps based on a Fully Convolutional Network (FCN) backbone, and produces the corresponding attention mask for each feature map. The learned attention masks can neatly highlight the regions of interest while suppress background clutter. Furthermore, we propose an efficient multi-scale feature interaction mechanism which is able to adaptively aggregate the hierarchical features. Based on the proposed mechanism, the features of different levels are adaptively re-weighted according to the local spatial structure and the surrounding contextual information. Consequently, the proposed modules are able to boost standard FCN architectures and result in an enhanced pixel-wise segmentation for street-level scene images. Extensive experiments on three public available street-level benchmarks demonstrate that the proposed Gated Attention Network (GANet) approach achieves consistently superior performance and outperforms the very recent state-of-the-art methods.

Ground and Multi-Class Classification of Airborne Laser Scanner Point Clouds Using Fully Convolutional Networks

Various classification methods have been developed to extract meaningful information from Airborne Laser Scanner (ALS) point clouds. However, the accuracy and the computational efficiency of the existing methods need to be improved, especially for the analysis of large datasets (e.g., at regional or national levels). In this paper, we present a novel deep learning approach to ground classification for Digital Terrain Model (DTM) extraction as well as for multi-class land-cover classification, delivering highly accurate classification results in a computationally efficient manner. Considering the top–down acquisition angle of ALS data, the point cloud is initially projected on the horizontal plane and converted into a multi-dimensional image. Then, classification techniques based on Fully Convolutional Networks (FCN) with dilated kernels are designed to perform pixel-wise image classification. Finally, labels are transferred from pixels to the original ALS points. We also designed a Multi-Scale FCN (MS-FCN) architecture to minimize the loss of information during the point-to-image conversion. In the ground classification experiment, we compared our method to a Convolutional Neural Network (CNN)-based method and LAStools software. We obtained a lower total error on both the International Society for Photogrammetry and Remote Sensing (ISPRS) filter test benchmark dataset and AHN-3 dataset in the Netherlands. In the multi-class classification experiment, our method resulted in higher precision and recall values compared to the traditional machine learning technique using Random Forest (RF); it accurately detected small buildings. The FCN achieved precision and recall values of 0.93 and 0.94 when RF obtained 0.91 and 0.92, respectively. Moreover, our strategy significantly improved the computational efficiency of state-of-the-art CNN-based methods, reducing the point-to-image conversion time from 47 h to 36 min in our experiments on the ISPRS filter test dataset. Misclassification errors remained in situations that were not included in the training dataset, such as large buildings and bridges, or contained noisy measurements.

Wavelet-enhanced convolutional neural network: a new idea in a deep learning paradigm.

Manual brain tumor segmentation is a challenging task that requires the use of machine learning techniques. One of the machine learning techniques that has been given much attention is the convolutional neural network (CNN). The performance of the CNN can be enhanced by combining other data analysis tools such as wavelet transform. In this study, one of the famous implementations of CNN, a fully convolutional network (FCN), was used in brain tumor segmentation and its architecture was enhanced by wavelet transform. In this combination, a wavelet transform was used as a complementary and enhancing tool for CNN in brain tumor segmentation. Comparing the performance of basic FCN architecture against the wavelet-enhanced form revealed a remarkable superiority of enhanced architecture in brain tumor segmentation tasks. Using mathematical functions and enhancing tools such as wavelet transform and other mathematical functions can improve the performance of CNN in any image processing task such as segmentation and classification.

A fully convolutional network for weed mapping of unmanned aerial vehicle (UAV) imagery.

Appropriate Site Specific Weed Management (SSWM) is crucial to ensure the crop yields. Within SSWM of large-scale area, remote sensing is a key technology to provide accurate weed distribution information. Compared with satellite and piloted aircraft remote sensing, unmanned aerial vehicle (UAV) is capable of capturing high spatial resolution imagery, which will provide more detailed information for weed mapping. The objective of this paper is to generate an accurate weed cover map based on UAV imagery. The UAV RGB imagery was collected in 2017 October over the rice field located in South China. The Fully Convolutional Network (FCN) method was proposed for weed mapping of the collected imagery. Transfer learning was used to improve generalization capability, and skip architecture was applied to increase the prediction accuracy. After that, the performance of FCN architecture was compared with Patch_based CNN algorithm and Pixel_based CNN method. Experimental results showed that our FCN method outperformed others, both in terms of accuracy and efficiency. The overall accuracy of the FCN approach was up to 0.935 and the accuracy for weed recognition was 0.883, which means that this algorithm is capable of generating accurate weed cover maps for the evaluated UAV imagery.

UFCN: a fully convolutional neural network for road extraction in RGB imagery acquired by remote sensing from an unmanned aerial vehicle

Road extraction in imagery acquired by low altitude remote sensing (LARS) carried out using an unmanned aerial vehicle (UAV) is presented. LARS is carried out using a fixed wing UAV with a high spatial resolution vision spectrum (RGB) camera as the payload. Deep learning techniques, particularly fully convolutional network (FCN), are adopted to extract roads by dense semantic segmentation. The proposed model, UFCN (U-shaped FCN) is an FCN architecture, which is comprised of a stack of convolutions followed by corresponding stack of mirrored deconvolutions with the usage of skip connections in between for preserving the local information. The limited dataset (76 images and their ground truths) is subjected to real-time data augmentation during training phase to increase the size effectively. Classification performance is evaluated using precision, recall, accuracy, F1 score, and brier score parameters. The performance is compared with support vector machine (SVM) classifier, a one-dimensional convolutional neural network (1D-CNN) model, and a standard two-dimensional CNN (2D-CNN). The UFCN model outperforms the SVM, 1D-CNN, and 2D-CNN models across all the performance parameters. Further, the prediction time of the proposed UFCN model is comparable with SVM, 1D-CNN, and 2D-CNN models. (C) 2018 Society of Photo-Optical Instrumentation Engineers (SPIE)

3-D Fully Convolutional Networks for Multimodal Isointense Infant Brain Image Segmentation.

Accurate segmentation of infant brain images into different regions of interest is one of the most important fundamental steps in studying early brain development. In the isointense phase (approximately 6-8 months of age), white matter and gray matter exhibit similar levels of intensities in magnetic resonance (MR) images, due to the ongoing myelination and maturation. This results in extremely low tissue contrast and thus makes tissue segmentation very challenging. Existing methods for tissue segmentation in this isointense phase usually employ patch-based sparse labeling on single modality. To address the challenge, we propose a novel 3-D multimodal fully convolutional network (FCN) architecture for segmentation of isointense phase brain MR images. Specifically, we extend the conventional FCN architectures from 2-D to 3-D, and, rather than directly using FCN, we intuitively integrate coarse (naturally high-resolution) and dense (highly semantic) feature maps to better model tiny tissue regions, in addition, we further propose a transformation module to better connect the aggregating layers; we also propose a fusion module to better serve the fusion of feature maps. We compare the performance of our approach with several baseline and state-of-the-art methods on two sets of isointense phase brain images. The comparison results show that our proposed 3-D multimodal FCN model outperforms all previous methods by a large margin in terms of segmentation accuracy. In addition, the proposed framework also achieves faster segmentation results compared to all other methods. Our experiments further demonstrate that: 1) carefully integrating coarse and dense feature maps can considerably improve the segmentation performance; 2) batch normalization can speed up the convergence of the networks, especially when hierarchical feature aggregations occur; and 3) integrating multimodal information can further boost the segmentation performance.

License plate detection based on fully convolutional networks

Fully convolutional networks (FCNs) have shown outstanding performance in image semantic segmentation, which is the key work in license plate detection (LPD). An FCN architecture for LPD is presented. First, a multiscale hierarchical network structure is used to combine multiscale and multilevel features produced by FCN. Then, an enhanced loss structure that contains three loss layers is defined to emphasize the license plates in images. Finally, the FCN generates prediction maps that directly show the location of license plates. Experiments show that our approach is more accurate than many state-of-the-art methods.

3D shape segmentation via shape fully convolutional networks

We propose a novel fully convolutional network architecture for shapes, denoted by Shape Fully Convolutional Networks (SFCN). 3D shapes are represented as graph structures in the SFCN architecture, based on novel graph convolution and pooling operations, which are similar to convolution and pooling operations used on images. Meanwhile, to build our SFCN architecture in the original image segmentation fully convolutional network (FCN) architecture, we also design and implement a generating operation with bridging function. This ensures that the convolution and pooling operation we have designed can be successfully applied in the original FCN architecture. In this paper, we also present a new shape segmentation approach based on SFCN. Furthermore, we allow more general and challenging input, such as mixed datasets of different categories of shapes which can prove the ability of our generalisation. In our approach, SFCNs are trained triangles-to-triangles by using three low-level geometric features as input. Finally, the feature voting-based multi-label graph cuts is adopted to optimise the segmentation results obtained by SFCN prediction. The experiment results show that our method can effectively learn and predict mixed shape datasets of either similar or different characteristics, and achieve excellent segmentation results.

Stacked fully convolutional networks with multi-channel learning: application to medical image segmentation

The automated segmentation of regions of interest (ROIs) in medical imaging is the fundamental requirement for the derivation of high-level semantics for image analysis in clinical decision support systems. Traditional segmentation approaches such as region-based depend heavily upon hand-crafted features and a priori knowledge of the user. As such, these methods are difficult to adopt within a clinical environment. Recently, methods based on fully convolutional networks (FCN) have achieved great success in the segmentation of general images. FCNs leverage a large labeled dataset to hierarchically learn the features that best correspond to the shallow appearance as well as the deep semantics of the images. However, when applied to medical images, FCNs usually produce coarse ROI detection and poor boundary definitions primarily due to the limited number of labeled training data and limited constraints of label agreement among neighboring similar pixels. In this paper, we propose a new stacked FCN architecture with multi-channel learning (SFCN-ML). We embed the FCN in a stacked architecture to learn the foreground ROI features and background non-ROI features separately and then integrate these different channels to produce the final segmentation result. In contrast to traditional FCN methods, our SFCN-ML architecture enables the visual attributes and semantics derived from both the fore- and background channels to be iteratively learned and inferred. We conducted extensive experiments on three public datasets with a variety of visual challenges. Our results show that our SFCN-ML is more effective and robust than a routine FCN and its variants, and other state-of-the-art methods.