Dense Conditional Random Field Research Articles

Weakly supervised semantic segmentation of leukocyte images based on class activation maps.

Leukocytes are an essential component of the human defense system, accurate segmentation of leukocyte images is a crucial step towards automating detection. Most existing methods for leukocyte images segmentation relied on fully supervised semantic segmentation (FSSS) with extensive pixel-level annotations, which are time-consuming and labor-intensive. To address this issue, this paper proposes a weakly supervised semantic segmentation (WSSS) approach for leukocyte images utilizing improved class activation maps (CAMs). Firstly, to alleviate ambiguous boundary problem between leukocytes and background, preprocessing technique is employed to enhance the image quality. Secondly, attention mechanism is added to refine the CAMs generated by improving the matching of local and global features. Random walks, dense conditional random fields and hole filling were leveraged to obtain final pseudo-segmentation labels. Finally, a fully supervised segmentation network is trained with pseudo-segmentation labels. The method is evaluated on BCCD and TMAMD datasets. Experimental results demonstrate that by employing the pseudo segmentation annotations generated through this method can be utilized to train UNet as close as possible to FSSS. This method effectively reduces manual annotation cost while achieving WSSS of leukocyte images.

Refined glacial lake extraction in a high-Asia region by deep neural network and superpixel-based conditional random field methods

Abstract. Remote sensing extraction of glacial lakes is an effective way of monitoring water body distribution and outburst events. At present, the lack of glacial lake datasets and the edge recognition problem of semantic segmentation networks lead to poor accuracy and inaccurate outlines of glacial lakes. Therefore, this study constructed a high-resolution dataset containing seven types of glacial lakes and proposed a refined glacial lake extraction method, which combines the LinkNet50 network for rough extraction and simple linear iterative clustering (SLIC) dense conditional random field (DenseCRF) for optimization. The results show that (1) with Google Earth images of 0.52 m resolution in the study area, the recall, precision, F1 score, and intersection over union (IoU) of glacial lake extraction based on the proposed method are 96.52 %, 92.49 %, 94.46 %, and 90.69 %, respectively, and (2) with the Google Earth images of 2.11 m resolution in the Qomolangma National Nature Reserve, 2300 glacial lakes with a total area of 65.17 km2 were detected by the proposed method. The area of the minimum glacial lake that can be extracted is 160 m2 (less than 6×6 pixels). This method has advantages in small glacial lake extraction and refined outline detection, which can be applied to extracting glacial lakes in the high-Asia region with high-resolution images.

OSC-CO2: coattention and cosegmentation framework for plant state change with multiple features.

Cosegmentation and coattention are extensions of traditional segmentation methods aimed at detecting a common object (or objects) in a group of images. Current cosegmentation and coattention methods are ineffective for objects, such as plants, that change their morphological state while being captured in different modalities and views. The Object State Change using Coattention-Cosegmentation (OSC-CO2) is an end-to-end unsupervised deep-learning framework that enhances traditional segmentation techniques, processing, analyzing, selecting, and combining suitable segmentation results that may contain most of our target object's pixels, and then displaying a final segmented image. The framework leverages coattention-based convolutional neural networks (CNNs) and cosegmentation-based dense Conditional Random Fields (CRFs) to address segmentation accuracy in high-dimensional plant imagery with evolving plant objects. The efficacy of OSC-CO2 is demonstrated using plant growth sequences imaged with infrared, visible, and fluorescence cameras in multiple views using a remote sensing, high-throughput phenotyping platform, and is evaluated using Jaccard index and precision measures. We also introduce CosegPP+, a dataset that is structured and can provide quantitative information on the efficacy of our framework. Results show that OSC-CO2 out performed state-of-the art segmentation and cosegmentation methods by improving segementation accuracy by 3% to 45%.

Extraction of building from remote sensing imagery base on multi-attention L-CAFSFM and MFFM

Building extraction from high-resolution remote sensing images is widely used in urban planning, land resource management, and other fields. However, the significant differences between categories in high-resolution images and the impact of imaging, such as atmospheric interference and lighting changes, make it difficult for high-resolution images to identify buildings. Therefore, detecting buildings from high-resolution remote sensing images is still challenging. In order to improve the accuracy of building extraction in high-resolution images, this paper proposes a building extraction method combining a bidirectional feature pyramid, location-channel attention feature serial fusion module (L-CAFSFM), and meticulous feature fusion module (MFFM). Firstly, richer and finer building features are extracted using the ResNeXt101 network and deformable convolution. L-CAFSFM combines feature maps from two adjacent levels and iteratively calculates them from high to low level, and from low to high level, to enhance the model’s feature extraction ability at different scales and levels. Then, MFFM fuses the outputs from the two directions to obtain building features with different orientations and semantics. Finally, a dense conditional random field (Dense CRF) improves the correlation between pixels in the output map. Our method’s precision, F-score, Recall, and IoU(Intersection over Union) on WHU Building datasets are 95.17%、94.83%、94.51% and 90.18%. Experimental results demonstrate that our proposed method has a more accurate effect in extracting building features from high-resolution image.

Deep-active-learning approach towards accurate right ventricular segmentation using a two-level uncertainty estimation.

The Right Ventricle (RV) is currently recognised to be a significant and important prognostic factor for various pathologies. Its assessment is made possible using Magnetic Resonance Imaging (CMRI) short-axis slices. Yet, due to the challenging issues of this cavity, radiologists still perform its delineation manually, which is tedious, laborious, and time-consuming. Therefore, to automatically tackle the RV segmentation issues, Deep-Learning (DL) techniques seem to be the axis of the most recent promising approaches. Along with its potential at dealing with shape variations, DL techniques highly requires a large number of labelled images to avoid over-fitting. Subsequently, with the produced large amounts of data in the medical industry, preparing annotated datasets manually is still time-consuming, and requires high skills to be accomplished. To benefit from a significant number of labelled and unlabelled CMRI images, a Deep-Active-Learning (DAL) approach is proposed in this paper to segment the RV. Thus, three main steps are distinguished. First, a personalised labelled dataset is gathered and augmented to allow initial learning. Secondly, a U-Net based architecture is modified towards efficient initial accuracy. Finally, a two-level uncertainty estimation technique is settled to enable the selection of complementary unlabelled data. The proposed pipeline is enhanced with a customised postprocessing step, in which epistemic uncertainty and Dense Conditional Random Fields are used. The proposed approach is tested on 791 images gathered from 32 public patients and 1230 images of 50 custom subjects. The obtained results show an increased dice coefficient from 0.86 to 0.91 with a decreased Hausdorff distance from 7.55 to 7.45.

Semantic Segmentation of Remote Sensing Imagery Based on Multiscale Deformable CNN and DenseCRF

The semantic segmentation of remote sensing images is a significant research direction in digital image processing. The complex background environment, irregular size and shape of objects, and similar appearance of different categories of remote sensing images have brought great challenges to remote sensing image segmentation tasks. Traditional convolutional-neural-network-based models often ignore spatial information in the feature extraction stage and pay less attention to global context information. However, spatial context information is important in complex remote sensing images, which means that the segmentation effect of traditional models needs to be improved. In addition, neural networks with a superior segmentation performance often suffer from the problem of high computational resource consumption. To address the above issues, this paper proposes a combination model of a modified multiscale deformable convolutional neural network (mmsDCNN) and dense conditional random field (DenseCRF). Firstly, we designed a lightweight multiscale deformable convolutional network (mmsDCNN) with a large receptive field to generate a preliminary prediction probability map at each pixel. The output of the mmsDCNN model is a coarse segmentation result map, which has the same size as the input image. In addition, the preliminary segmentation result map contains rich multiscale features. Then, the multi-level DenseCRF model based on the superpixel level and the pixel level is proposed, which can make full use of the context information of the image at different levels and further optimize the rough segmentation result of mmsDCNN. To be specific, we converted the pixel-level preliminary probability map into a superpixel-level predicted probability map according to the simple linear iterative clustering (SILC) algorithm and defined the potential function of the DenseCRF model based on this. Furthermore, we added the pixel-level potential function constraint term to the superpixel-based Gaussian potential function to obtain a combined Gaussian potential function, which enabled our model to consider the features of various scales and prevent poor superpixel segmentation results from affecting the final result. To restore the contour of the object more clearly, we utilized the Sketch token edge detection algorithm to extract the edge contour features of the image and fused them into the potential function of the DenseCRF model. Finally, extensive experiments on the Potsdam and Vaihingen datasets demonstrated that the proposed model exhibited significant advantages compared to the current state-of-the-art models.

Deep supervision and atrous inception-based U-Net combining CRF for automatic liver segmentation from CT

Due to low contrast and the blurred boundary between liver tissue and neighboring organs sharing similar intensity values, the problem of liver segmentation from CT images has not yet achieved satisfactory performance and remains a challenge. To alleviate these problems, we introduce deep supervision (DS) and atrous inception (AI) technologies with conditional random field (CRF) and propose three major improvements that are experimentally shown to have substantive and practical value. First, we replace the encoder's standard convolution with the residual block. Residual blocks can increase the depth of the network. Second, we provide an AI module to connect the encoder and decoder. AI allows us to obtain multi-scale features. Third, we incorporate the DS mechanism into the decoder. This helps to make full use of information of the shallow layers. In addition, we employ the Tversky loss function to balance the segmented and non-segmented regions and perform further refinement with a dense CRF. Finally, we extensively validate the proposed method on three public databases: LiTS17, 3DIRCADb, and SLiver07. Compared to the state-of-the-art methods, the proposed method achieved increased segmentation accuracy for the livers with low contrast and the fuzzy boundary between liver tissue and neighboring organs and is, therefore, more suited for automatic segmentation of these livers.

Dynamic dense CRF inference for video segmentation and semantic SLAM

The dense conditional random field (dense CRF) is an effective post-processing tool for image/video segmentation and semantic SLAM. In this paper, we extend the traditional dense CRF inference algorithm to incremental sensor data modelling. The algorithm efficiently infers the maximum a posteriori probability (MAP) solution for a dynamically changing dense CRF model that is applied to incremental multi-class video segmentation and semantic SLAM. The computational cost is roughly proportional to the total change in the Gaussian pairwise edges of the dense CRF. In our system, with an increase in the number of frames of the sensor data, MAP calculations take approximately the same time to compute the overall three-dimensional dense CRF modelled for the entire video. Compared with the traditional dense CRF for video segmentation, this method is more suitable for incremental (in-line) video segmentation and robot semantic SLAM. The results of experiments show that if part of a pairwise edge is altered, our dynamic algorithm is significantly faster than the widely known standard dense CRF algorithm. In addition, the accuracy of its inference does not change. Several multi-class video segmentation tests confirmed the efficiency of inference of the algorithm. In another application, we used the dynamic dense CRF to incrementally integrate robot SLAM and video segmentation. The results show that an accurate SLAM can improve the accuracy of video segmentation, and the computational cost of the dense CRF MAP can be constrained over a constant range. The application of our algorithm is not limited to video segmentation: It is generic, and can be used to yield similar improvements in many optimization solutions for MAP in dynamically changing models.

A Land Use Classification Model Based on Conditional Random Fields and Attention Mechanism Convolutional Networks

Land use is used to reflect the expression of human activities in space, and land use classification is a way to obtain accurate land use information. Obtaining high-precision land use classification from remote sensing images remains a significant challenge. Traditional machine learning methods and image semantic segmentation models are unable to make full use of the spatial and contextual information of images. This results in land use classification that does not meet high-precision requirements. In order to improve the accuracy of land use classification, we propose a land use classification model, called DADNet-CRFs, that integrates an attention mechanism and conditional random fields (CRFs). The model is divided into two modules: the Dual Attention Dense Network (DADNet) and CRFs. First, the convolution method in the UNet network is modified to Dense Convolution, and the band-hole pyramid pooling module, spatial location attention mechanism module, and channel attention mechanism module are fused at appropriate locations in the network, which together form DADNet. Second, the DADNet segmentation results are used as a priori conditions to guide the training of CRFs. The model is tested with the GID dataset, and the results show that the overall accuracy of land use classification obtained with this model is 7.36% and 1.61% higher than FCN-8s and BiSeNet in classification accuracy, 11.95% and 1.81% higher in MIoU accuracy, and with a 9.35% and 2.07% higher kappa coefficient, respectively. The proposed DADNet-CRFs model can fully use the spatial and contextual semantic information of high-resolution remote sensing images, and it effectively improves the accuracy of land use classification. The model can serve as a highly accurate automatic classification tool for land use classification and mapping high-resolution images.

Solar Filament Detection Based on Improved DeepLab V3+

A novel solar filament detection method based on an improved DeepLab V3+ is proposed to address the low detection accuracy of small solar filaments in Hα full-disk solar images. First, the Xception structure of the backbone network is fine-tuned, and the low-level feature information of the filaments is added to the decoder module of the network to improve the utilization of the solar filament features. Second, the receptive field of dilated convolution is expanded, and the information utilization rate is increased via cascaded dilated convolution to improve the detection accuracy of the small solar filaments. In the decoder module, two depthwise separable convolutions are used instead of ordinary convolutions to reduce incomplete detections. Finally, a dense conditional random field is added to optimize the edge of the detection results. Experiments on a public data set comprising full-disk Hα images show that compared with the original Deeplab V3+ algorithm, the proposed method improves the mean pixel accuracy, mean intersection over union, and F1-score by 1.86%, 1.95%, and 2.18%, respectively, which also demonstrates its superiority over other existing solar filament detection algorithms.

Robust Multiseasonal Ice Classification From High-Resolution X-Band SAR

Automated solutions for sea ice-type classification from synthetic aperture radar (SAR) imagery offer an opportunity to monitor sea ice, unimpeded by cloud cover or the arctic night. However, there is a common struggle to obtain accurate classifications year round, particularly in the melt and freeze-up seasons. During these seasons, the radar backscatter signal is affected by wet snow cover, obscuring information about underlying ice types. By using additional spatiotemporal contextual data and a combination of convolutional neural networks and a dense conditional random field, we can mitigate these problems and obtain a single classifier that is able to classify accurately at 3.5-m spatial resolution for five different classes of sea ice surface from October to May. During the near year-long drift of the Multidisciplinary Drifting Observatory for the Study of the Arctic Climate (MOSAiC) expedition, we collected satellite scenes of the same patch of Arctic pack ice with X-band SAR with a revisit time of less than a day on average. Combined with <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">in situ</i> observations of the local ice properties, this offers up the unprecedented opportunity to perform a detailed and quantitative assessment of the robustness of our classifier for level, deformed, and heavily deformed ice. For these three classes, we can perform accurate classification with a probability >95% and calculate a lower bound for the robustness between 85% and 88%.

Research on a Segmentation Algorithm for the Tujia Brocade Images Based on Unsupervised Gaussian Mixture Clustering.

Tujia brocades are important carriers of Chinese Tujia national culture and art. It records the most detailed and real cultural history of Tujia nationality and is one of the National Intangible Cultural Heritage. Classic graphic elements are separated from Tujia brocade patterns to establish the Tujia brocade graphic element database, which is used for the protection and inheritance of traditional national culture. Tujia brocade dataset collected a total of more than 200 clear Tujia brocade patterns and was divided into seven categories, according to traditional meanings. The weave texture of a Tujia brocade is coarse, and the textural features of the background are obvious, so classical segmentation algorithms cannot achieve good segmentation effects. At the same time, deep learning technology cannot be used because there is no standard Tujia brocade dataset. Based on the above problems, this study proposes a method based on an unsupervised clustering algorithm for the segmentation of Tujia brocades. First, the cluster number K is calculated by fusing local binary patterns (LBP) and gray-level co-occurrence matrix (GLCM) characteristic values. Second, clustering and segmentation are conducted on each input Tujia brocade image by adopting a Gaussian mixture model (GMM) to obtain a preliminary segmentation image, wherein the image yielded after preliminary segmentation is rough. Then, a method based on voting optimization and dense conditional random field (DenseCRF) (CRF denotes conditional random filtering) is adopted to optimize the image after preliminary segmentation and obtain the image segmentation results. Finally, the desired graphic element contour is extracted through interactive cutting. The contributions of this study include: (1) a calculation method for the cluster number K wherein the experimental results show that the effect of the clustering number K chosen in this paper is ideal; (2) an optimization method for the noise points of Tujia brocade patterns based on voting, which can effectively eliminate isolated noise points from brocade patterns.

Invariant descriptors for intrinsic reflectance optimization.

Intrinsic image decomposition aims to factorize an image into albedo (reflectance) and shading (illumination) sub-components. Being ill posed and under-constrained, it is a very challenging computer vision problem. There are infinite pairs of reflectance and shading images that can reconstruct the same input. To address the problem, Intrinsic Images in the Wild by Bell et al. provides an optimization framework based on a dense conditional random field (CRF) formulation that considers long-range material relations. We improve upon their model by introducing illumination invariant image descriptors: color ratios. The color ratios and the intrinsic reflectance are both invariant to illumination and thus are highly correlated. Through detailed experiments, we provide ways to inject the color ratios into the dense CRF optimization. Our approach is physics based and learning free and leads to more accurate and robust reflectance decompositions.

Semantic Point Cloud Segmentation Using Fast Deep Neural Network and DCRF.

Accurate segmentation of entity categories is the critical step for 3D scene understanding. This paper presents a fast deep neural network model with Dense Conditional Random Field (DCRF) as a post-processing method, which can perform accurate semantic segmentation for 3D point cloud scene. On this basis, a compact but flexible framework is introduced for performing segmentation to the semantics of point clouds concurrently, contribute to more precise segmentation. Moreover, based on semantics labels, a novel DCRF model is elaborated to refine the result of segmentation. Besides, without any sacrifice to accuracy, we apply optimization to the original data of the point cloud, allowing the network to handle fewer data. In the experiment, our proposed method is conducted comprehensively through four evaluation indicators, proving the superiority of our method.

Monitoring Urban Deprived Areas with Remote Sensing and Machine Learning in Case of Disaster Recovery

Rapid urbanization and increasing population in cities with a large portion of them settled in deprived neighborhoods, mostly defined as slum areas, have escalated inequality and vulnerability to natural disasters. As a result, monitoring such areas is essential to provide information and support decision-makers and urban planners, especially in case of disaster recovery. Here, we developed an approach to monitor the urban deprived areas over a four-year period after super Typhoon Haiyan, which struck Tacloban city, in the Philippines, in 2013, using high-resolution satellite images and machine learning methods. A Support Vector Machine classification method supported by a local binary patterns feature extraction model was initially performed to detect slum areas in the pre-disaster, just after/event, and post-disaster images. Afterward, a dense conditional random fields model was employed to produce the final slum areas maps. The developed method detected slum areas with accuracies over 83%. We produced the damage and recovery maps based on change analysis over the detected slum areas. The results revealed that most of the slum areas were reconstructed 4 years after Typhoon Haiyan, and thus, the city returned to the pre-existing vulnerability level.

A hybrid convolutional architecture for accurate image manipulation localization at the pixel-level

Advanced image processing techniques can easily edit images without leaving any visible traces, making manipulation detection and localization for forensics analysis a challenging task. Few studies can simultaneously locate tampered objects accurately and refine contours of tampered regions effectively. In this study, we propose an effective and novel hybrid architecture, named Pixel-level Image Tampering Localization Architecture (PITLArc), which integrates the advantages of top-down detection-based methods and bottom-up segmentation-based methods. Moreover, we provide a typical fusion implementation of our proposed hybrid architecture on one outstanding detection-based method (two-stream faster region-based convolutional neural network (RGB-N)) and two segmentation-based methods (Multi-Scale Convolution Neural Networks (MSCNNs) and Dual-domain Convolutional Neural Networks (DCNNs)) to evaluate the effectiveness of the proposed architecture. The three methods can be integrated into our proposed PITLArc to significantly improve their performance. Other detection and segmentation algorithms (not limited to the three aforementioned methods) can also be integrated into our architecture to improve their performance. Moreover, a Dense Conditional Random Fields (DenseCRFs)-based post-processing method is introduced to further optimize the details of tampered regions. Experiments validate the effectiveness of the proposed architecture.

A Deep Extractor for Visual Rail Surface Inspection

Rail surface inspection using a visual inspection system is an important task in the maintenance of railway networks. Recent inspection approaches have attempted to harness low-level features to identify visual defects. These methods' main issues are the limitations of low-level features and prior information about rail surfaces that show a wide variety of visual appearances on dynamic backgrounds. To overcome these problems, we propose a deep extractor (DE) that combines the strengths of fully convolutional networks and conditional random fields (CRFs), so that the deep network can learn the abstract features needed for the specified inspection task. Specifically, motivated by fully convolutional networks, a bilateral fully convolutional network containing two branches is proposed here: one encode-decode branch focuses semantic meaning; the other decode-encode branch encodes information on tiny defects. The two branches are aggregated to obtain a high-level feature map. Furthermore, the mean-field inference of a dense CRF with Gaussian pairwise potentials formulated as a recurrent neural network (CRF-RNN) is introduced to contribute to smoothing constraints and obtaining a fine-grained inspection result. By doing so, it can avoid offline post-processing and further lead the whole deep architecture to achieve end-to-end deep learning. Compared with classical approaches, our approach outperforms state-of-art approaches in analyzing two publicly available datasets.

Deeply supervised U‐Net for mass segmentation in digital mammograms

AbstractMass detection is a critical process in the examination of mammograms. The shape and texture of the mass are key parameters used in the diagnosis of breast cancer. To recover the shape of the mass, semantic segmentation is found to be more useful rather than mere object detection (or) localization. The main challenges involved in the mass segmentation include: (a) low signal to noise ratio (b) indiscernible mass boundaries, and (c) more false positives. These problems arise due to the significant overlap in the intensities of both the normal parenchymal region and the mass region. To address these challenges, deeply supervised U‐Net model (DS U‐Net) coupled with dense conditional random fields (CRFs) is proposed. Here, the input images are preprocessed using CLAHE and a modified encoder‐decoder‐based deep learning model is used for segmentation. In general, the encoder captures the textual information of various regions in an input image, whereas the decoder recovers the spatial location of the desired region of interest. The encoder‐decoder‐based models lack the ability to recover the non‐conspicuous and spiculated mass boundaries. In the proposed work, deep supervision is integrated with a popular encoder‐decoder model (U‐Net) to improve the attention of the network toward the boundary of the suspicious regions. The final segmentation map is also created as a linear combination of the intermediate feature maps and the output feature map. The dense CRF is then used to fine‐tune the segmentation map for the recovery of definite edges. The DS U‐Net with dense CRF is evaluated on two publicly available benchmark datasets CBIS‐DDSM and INBREAST. It provides a dice score of 82.9% for CBIS‐DDSM and 79% for INBREAST.

Refined UNet V2: End-to-End Patch-Wise Network for Noise-Free Cloud and Shadow Segmentation

Cloud and shadow detection is an essential prerequisite for further remote sensing processing, whereas edge-precise segmentation remains a challenging issue. In Refined UNet, we considered the aforementioned task and proposed a two-stage pipeline to achieve the edge-precise segmentation. The isolated segmentation regions in Refined UNet, however, bring inferior visualization and should be sufficiently eliminated. Moreover, an end-to-end model is also expected to jointly predict and refine the segmentation results. In this paper, we propose the end-to-end Refined UNet v2 to achieve joint prediction and refinement of cloud and shadow segmentation, which is capable of visually neutralizing redundant segmentation pixels or regions. To this end, we inherit the pipeline of Refine UNet, revisit the bilateral message passing in the inference of conditional random field (CRF), and then develop a novel bilateral strategy derived from the Guided Gaussian filter. Derived from a local linear model of denoising, our v2 can considerably remove isolated segmentation pixels or regions, which is able to yield “cleaner” results. Compared to the high-dimensional Gaussian filter, the Guided Gaussian filter-based message-passing strategy is quite straightforward and easy to implement so that a brute-force implementation can be easily given in GPU frameworks, which is potentially efficient and facilitates embedding. Moreover, we prove that Guided Gaussian filter-based message passing is highly relevant to the Gaussian bilateral term in Dense CRF. Experiments and results demonstrate that our v2 is quantitatively comparable to Refined UNet, but can visually outperform that from the noise-free segmentation perspective. The comparison of time consumption also supports the potential efficiency of our v2.

Beyond the Patchwise Classification: Spectral-Spatial Fully Convolutional Networks for Hyperspectral Image Classification

In recent years, patchwise classification methods are commonly adopted when dealing with the hyperspectral image (HSI) classification. Despite their promising results from the perspective of accuracy, the efficiency of these methods can hardly be ensured since there are redundant computations between adjacent patches. In this paper, we propose a spectral-spatial fully convolutional network for HSI classification with an end-to-end, pixel-to-pixel architecture. Compared with patchwise methods, the proposed framework can avoid the patch extraction and is more efficient. Since the training samples in HSIs are highly sparse, the training strategy in original fully convolutional networks is no longer feasible for HSIs. To solve this problem, we propose a novel mask matrix to assist the back-propagation in the training stage. Considering the importance of spectral and spatial features may vary for different objects and scenes, we combine both features with two weighting factors which can be adaptively learned during the network training. Besides, the dense conditional random field (CRF) is introduced into the framework to further balance the local and global information. Experiments on three benchmark HSI data sets demonstrate that the proposed method can yield competitive results with less time costs compared with patchwise methods.