Top PDF A Classification Algorithm for Hyperspectral Images based on Synergetics Theory

A Classification Algorithm for Hyperspectral Images based on Synergetics Theory

A Classification Algorithm for Hyperspectral Images based on Synergetics Theory

A Classification Algorithm for Hyperspectral Images Based on Synergetics Theory Daniele Cerra, Member, IEEE, Rupert Müller, and Peter Reinartz Abstract—This paper presents a classification methodology for hyperspectral data based on synergetics theory. Pattern recogni- tion algorithms based on synergetics have been applied to images in the spatial domain with limited success in the past, given their dependence on the rotation, shifting, and scaling of the images. These drawbacks can be discarded if such methods are applied to data acquired by a hyperspectral sensor in the spectral domain, as each single spectrum, related to an image element in the hy- perspectral scene, can be analyzed independently. The spectrum is first projected in a space spanned by a set of user-defined pro- totype vectors, which belong to some classes of interest, and then attracted by a final state associated to a prototype. The spectrum can thus be classified, establishing a first attempt at performing a pixel-wise image classification using notions derived from syner- getics. As typical synergetics-based systems have the drawback of a rigid training step, we introduce a new procedure which allows the selection of a training area for each class of interest, used to weight the prototype vectors through attention parameters and to produce a more accurate classification map through plurality vote of independent classifications. As each classification is in principle obtained on the basis of a single training sample per class, the proposed technique could be particularly effective in tasks where only a small training data set is available. The results presented are promising and often outperform state-of-the-art classification methodologies, both general and specific to hyperspectral data.
Show more

12 Read more

Dimension Reduction and Classification of Hyperspectral Images based on Neural Network Sensitivity Analysis and Multi-instance Learning

Dimension Reduction and Classification of Hyperspectral Images based on Neural Network Sensitivity Analysis and Multi-instance Learning

Abstract. Hyperspectral remote image sensing is a rapidly developing integrated technology used widely in numerous areas. The rich spectral information from hyperspectral images aids in recognition and classification of many types of objects, but the high dimensionality of these images leads to information redundancy. In this paper, we used sensitivity analysis for dimension reduction. However, another challenge is that hyperspectral images identify objects as either a "different body with the same spectrum" or "same body with a different spectrum." Therefore, it is difficult to maintain the correct correspondence between ground objects and samples, which hinders classification of the images. This issue can be addressed using multi-instance learning for classification. In our proposed method, we combined neural network sensitivity analysis with a multi- instance learning algorithm based on a support vector machine to achieve dimension reduction and accurate classification for hyperspectral images. Experimental results demonstrated that our method provided strong overall classification effectiveness when compared with prior methods.
Show more

26 Read more

Hyperspectral Image Classification  Based on Hierarchical SVM  Algorithm for Improving  Overall Accuracy

Hyperspectral Image Classification Based on Hierarchical SVM Algorithm for Improving Overall Accuracy

HSI classification is a significant challenge in remote sensing applications. Generally, the HIS classification algorithms fall into three categories: supervised, unsupervised, and semi-supervised. Due to the high feature space dimension of the hyperspectral images, the supervised algorithms are encountered with the Hughes phenomenon. Two approaches are proposed for solving this problem. The first, the semi-supervised algorithm [1] prevent from Hughes phenomenon with predicting initial labels for the test pixels. The feature space reduction [2] which includes two different methods, feature extraction [3] and feature selec- tion [4] is the second approach for reducing computational complexity and in- creasing prediction accuracy. In [5], a Genetic Algorithm (GA) based wrapper method is presented for classification of hyperspectral image using (SVM), a state-of-art classifier that has found success in a variety of areas.
Show more

11 Read more

Gravitation-based edge detection for hyperspectral images

Gravitation-based edge detection for hyperspectral images

Abstract: Edge detection is one of the key issues in the field of computer vision and remote sensing image analysis. Although many different edge-detection methods have been proposed for gray-scale, color, and multispectral images, they still face difficulties when extracting edge features from hyperspectral images (HSIs) that contain a large number of bands with very narrow gap in the spectral domain. Inspired by the clustering characteristic of the gravitational theory, a novel edge-detection algorithm for HSIs is presented in this paper. In the proposed method, we first construct a joint feature space by combining the spatial and spectral features. Each pixel of HSI is assumed to be a celestial object in the joint feature space, which exerts gravitational force to each of its neighboring pixel. Accordingly, each object travels in the joint feature space until it reaches a stable equilibrium. At the equilibrium, the image is smoothed and the edges are enhanced, where the edge pixels can be easily distinguished by calculating the gravitational potential energy. The proposed edge-detection method is tested on several benchmark HSIs and the obtained results were compared with those of four state-of-the-art approaches. The experimental results confirm the efficacy of the proposed method. Keywords: edge detection; hyperspectral image; gravitation; remote sensing; feature space
Show more

23 Read more

Survey on Region Growing Segmentation and Classification for Hyperspectral Images

Survey on Region Growing Segmentation and Classification for Hyperspectral Images

This is similar to the marker controlled segmentation approach, this algorithm works out to generate classification derived markers and do segmentation and classification processing based on the markers. An improvement is that instead of using a gradient, the image and constructing a minimum spanning forest (MSF) where markers are the roots. A polling technique is implemented to identify connected components which work together with the minimum spanning forest to generate optimal segmentation and classification map. A demerit of this method is that if no marker is chosen for a particular spatial structure, this sptial region will be lost in the final classification map.
Show more

6 Read more

Spatial-spectral classification of hyperspectral images : a deep learning framework with Markov random fields based modeling

Spatial-spectral classification of hyperspectral images : a deep learning framework with Markov random fields based modeling

Abstract: For spatial-spectral classification of hyperspectral images (HSI), a deep learning framework is proposed in this paper, which consists of convolutional neural networks (CNN) and Markov random fields (MRF). Firstly, a CNN model to learn the deep spectral feature from the HSI is built and the class posterior probability distribution is estimated. The CNN with a dropout layer can relieve the overfitting in classification. The CNN is utilized as a pixel-classifier, so it only works in the spectral domain. Then, the spatial information will be encoded by MRF-based multilevel logistic (MLL) prior for regularizing the classification. To derive the correlation of both spectral and spatial features for improving algorithm performance, the marginal probability distribution in HSI is learned using MRF-based loopy belief propagation (LBP). In comparison with several state-of-the-art approaches for data classification on 3 publicly available HSI datasets, experimental results have demonstrated the superior performance of the proposed methodology.
Show more

12 Read more

Extended Self-Dual Attribute Profiles for the Classification of Hyperspectral Images

Extended Self-Dual Attribute Profiles for the Classification of Hyperspectral Images

methods, which might face significant challenges especially when dealing with data of high dimensionality [11]. Data mining techniques are included in the classification of remote sensing images for two main reasons: to counteract the Hughes phenomenon [12] (i.e., the combination of high dimensional data and a small number of training samples) and to reduce the required computational load. Data mining techniques present in the literature include supervised and unsupervised, paramet- ric and nonparametric, linear and nonlinear methods, which all seek to identify the relevant informative subspace. Indeed, the original feature space may not be the most effective domain for representing hyperspectral data. This work adopts the Non- parametric Weighted Feature Extraction (NWFE) technique which is an efficient algorithm for high dimensional multi- class pattern recognition problems [10]. NWFE is a supervised method (i.e., it computes the transformation according to the properties of the training set), and assigns different weights to each sample in order to compute the weighted means. Since NWFE is based on a nonparametric extension of scatter matrices (i.e., between-class and within-class), the algorithm is able to extract a desired number of features (higher than the number of classes) and can work well even for data that are not in Gaussian distribution [10].
Show more

6 Read more

An Improved Segmentation Algorithm For X-Ray Images Based On Adaptive Thresholding Classification

An Improved Segmentation Algorithm For X-Ray Images Based On Adaptive Thresholding Classification

In Medical Decision Support System (MDSS), an X-Ray image analysis is a sophisticated process in which an X-Ray image of human body is passed as an input and it generates output to support decision of medical practitioners. Many researchers have shown their interest in segmentation of biomedical images. If edge detection and segmentation are performed accurately and precisely, further processing on X-Ray image becomes much easier to generate interpretation reports. Following paragraphs discuss about biomedical image segmentation related work of some latest methods/algorithms.Authors of paper [9] have presented a novel segmentation technique based on MsGKFCM clustering which automatically segments the breast US (ultra sound) images. This method achieves Accuracy (Mean±STD) 97.3158±0.0409, F-score (Mean±STD) 92.310±0.0246, Specificity (Mean±STD) 12.7777±0.0246, Precision (Mean±STD) 86.4956±0.0244, Recall (Mean±STD) 94.3291± 0.0385. This method can be further improved by taking more than 90 image dataset. Researchers of paper [10] have proposed a novel region-growing segmentation method namely IPKD based on possibilistic theory on synthetic images as well as mammographic images from MIAS database. The advantage of this mechanism is to provide iterative diffusion of per-pixel certain knowledge to surrounding pixels in order to progressively refine the segmentation process. IPKD‘s performance (in terms of recognition rate, 94.37% and global predictive rate, 92.18%) is compared with three relevant reference methods: level-set, Fuzzy C-Mean and region growing methods. The IPKD approach outperforms the other three methods, respectively, at the recognition rates of
Show more

7 Read more

Advanced Techniques for the Classification of Very High Resolution and Hyperspectral Remote Sensing Images

Advanced Techniques for the Classification of Very High Resolution and Hyperspectral Remote Sensing Images

d) A novel protocol for accuracy assessment in classification of very high resolution images With the availability of VHR images acquired by satellite multispectral scanners, it is possi- ble to acquire detailed information on the shape and the geometry of the objects present on the ground. This detailed information can be exploited by automatic classification systems to gener- ate land-cover maps that exhibit a high degree of geometrical details. The precision that the clas- sification system can afford in the characterization of the geometrical properties of the objects present on the ground is particularly relevant in many practical applications, e.g., in urban area mapping, building characterization, target detection, crop fields classification in precision farm- ing, etc. In this context, a major open issue in classification of VHR images is the lack of ade- quate strategies for a precise evaluation of the quality of the produced thematic maps. The most common accuracy assessment methodology in classification of VHR images is based on the computation of thematic accuracy measures according to collected reference data. However, the thematic accuracy alone does not result sufficient for effectively characterizing the geometrical properties of the objects recognized in a map, because it assesses the correctness of the land- cover labels of sparse test pixels (or regions of interests) that do not model the actual shape of the objects in the scene. Thus, often maps derived by different classifiers (or with different pa- rameter values for the same classifier) that have similar thematic accuracy exhibit significantly different geometric properties (and thus global quality). For this reason, in many real classifica- tion problems the quality of the maps obtained by the classification of VHR data is assessed also through a visual inspection. However, this procedure can provide just a subjective evaluation of the map quality that can not be quantified. Thus, it is important to develop accuracy assessment protocols for a precise, objective, and quantitative characterization of the quality of thematic maps in terms of both thematic and geometric properties. These protocols could be used not only for assessing the quality of thematic maps generated by different classification systems, but also for better driving the model selection of a single classifier, i.e., the selection of the optimum values for the free parameters of a supervised categorization algorithm.
Show more

158 Read more

Spatial-spectral classification of hyperspectral images : a deep learning framework with Markov random fields based modeling

Spatial-spectral classification of hyperspectral images : a deep learning framework with Markov random fields based modeling

Abstract: For spatial-spectral classification of hyperspectral images (HSI), a deep learning framework is proposed in this paper, which consists of convolutional neural networks (CNN) and Markov random fields (MRF). Firstly, a CNN model to learn the deep spectral feature from the HSI is built and the class posterior probability distribution is estimated. The CNN with a dropout layer can relieve the overfitting in classification. The CNN is utilized as a pixel-classifier, so it only works in the spectral domain. Then, the spatial information will be encoded by MRF-based multilevel logistic (MLL) prior for regularizing the classification. To derive the correlation of both spectral and spatial features for improving algorithm performance, the marginal probability distribution in HSI is learned using MRF-based loopy belief propagation (LBP). In comparison with several state-of-the-art approaches for data classification on 3 publicly available HSI datasets, experimental results have demonstrated the superior performance of the proposed methodology.
Show more

13 Read more

Developmentof a classification algorithm for efficient handling of multiple classes in sorting systems based on hyperspectral imaging

Developmentof a classification algorithm for efficient handling of multiple classes in sorting systems based on hyperspectral imaging

When dealing with practical applications of hyperspectral imaging, the development of efficient, fast and flexible classification algorithms is of the utmost importance. Indeed, the optimal classification method should be able, in a reasonable time, to maximise the separation between the classes of interest and, at the same time, to correctly reject possible outlier samples. To this aim, a new extension of Partial Least Squares Discriminant Analysis (PLS-DA), namely Soft PLS-DA, has been implemented. The basic engine of Soft PLS-DA is the same as PLS-DA, but class assignment is subjected to some additional criteria which allow samples not belonging to the target classes to be identified and rejected. The proposed approach was tested on a real case study of plastic waste sorting based on near infrared hyperspectral imaging. Household plastic waste objects made of the six recyclable plastic polymers commonly used for packaging were collected and imaged using a hyperspectral camera mounted on an indus- trial sorting system. In addition, paper and not recyclable plastics were also considered as potential foreign materials that are commonly found in plastic waste. For classification purposes, the Soft PLS-DA algorithm was integrated into a hierarchical classification tree for the discrimination of the different plastic polymers. Furthermore, Soft PLS-DA was also coupled with sparse-based variable selection to identify the relevant variables involved in the classification and to speed up the sorting process. The tree-structured classification model was successfully validated both on a test set of representative spectra of each material for a quantitative evaluation, and at the pixel level on a set of hyperspectral images for a qualitative assessment.
Show more

15 Read more

Varietal classification of rice seeds using RGB and hyperspectral images

Varietal classification of rice seeds using RGB and hyperspectral images

ABSTRACT Inspection of rice seeds is a crucial task for plant nurseries and farmers since it ensures seed quality when growing seedlings. Conventionally, this process is performed by expert inspectors who manually screen large samples of rice seeds to identify their species and assess the cleanness of the batch. In the quest to automate the screening process through machine vision, a variety of approaches utilise appearance-based features extracted from RGB images while others utilise the spectral information acquired using Hyperspectral Imaging (HSI) systems. Most of the literature on this topic benchmarks the performance of new discrimination models using only a small number of species. Hence, it is unclear whether or not model performance variance confirms the effectiveness of proposed algorithms and features, or if it can be simply attributed to the inter-class/intra-class variations of the dataset itself. In this paper, a novel method to automatically screen and classify rice seed samples is proposed using a combination of spatial and spectral features, extracted from high resolution RGB and hyperspectral images. The proposed system is evaluated using a large dataset of 8,640 rice seeds sampled from a variety of 90 different species. The dataset is made publicly available to facilitate robust comparison and benchmarking of other existing and newly proposed techniques going forward. The proposed algorithm is evaluated on this large dataset and the experimental results show the effectiveness of the algorithm to eliminate impure species by combining spatial features extracted from high spatial resolution images and spectral features from hyperspectral data cubes.
Show more

12 Read more

RVM Classification of Hyperspectral Images Based on Wavelet Kernel Non-negative Matrix Fractorization

RVM Classification of Hyperspectral Images Based on Wavelet Kernel Non-negative Matrix Fractorization

A novel kernel framework for hyperspectral image classification based on relevance vector machine (RVM) is presented in this paper. The new feature extraction algorithm based on Mexican hat wavelet kernel non-negative matrix factorization (WKNMF) for hyperspectral remote sensing images is proposed. By using the feature of multi-resolution analysis, the new method of nonlinear mapping capability based on kernel NMF can be improved. The new classification framework of hyperspectral image data combined with the novel WKNMF and RVM. The simulation experimental results on HYDICE and AVIRIS data sets are both show that the classification accuracy of proposed method compared with other experiment methods even can be improved over 10% in some cases and the classification precision of small sample data area can be improved effectively.
Show more

9 Read more

A new Bayesian unmixing algorithm for hyperspectral images mitigating endmember variability

A new Bayesian unmixing algorithm for hyperspectral images mitigating endmember variability

This paper presents an unsupervised Bayesian algorithm for hyperspectral image unmixing accounting for endmember variability. Each image pixel is modeled by a linear combina- tion of random endmembers to take into account endmember variability in the image. The coefficients of this linear com- bination (referred to as abundances) allow the proportions of each material (endmembers) to be quantified in the image pixel. An additive noise is also considered in the proposed model generalizing the normal compositional model. The proposed Bayesian algorithm exploits spatial correlations between adjacent pixels of the image and provides spectral information by achieving a spectral unmixing. It estimates both the mean and the covariance matrix of each endmember in the image. A spatial classification is also obtained based on the estimated abundances. Simulations conducted with synthetic and real data show the potential of the proposed model and the unmixing performance for the analysis of hyperspectral images.
Show more

6 Read more

Unsupervised Classification of Hyperspectral Images based on Spectral Features

Unsupervised Classification of Hyperspectral Images based on Spectral Features

1.7 Thesis Outline This thesis contains four main chapters each consisting of sections. Chapter 2: Theory of Basic Methods In this chapter, approach to classification, basic theory of PCA Data extraction Algorithm, basic classification Algorithms like K-Means, K-Nearest Neighbor are explained. With the basic algorithms that are going to be used here are cleared in the this chapter, I am going to introduce some new algorithms some of which are specifically for the task of classification of land cover in Hyperspectral images and one is a general classifier technique in the next chapter onwards.
Show more

49 Read more

Frontiers in Spectral-Spatial Classification of Hyperspectral Images

Frontiers in Spectral-Spatial Classification of Hyperspectral Images

The main family of probabilistic graphical models that have been extensively applied to HSI classification is given by Markov random fields (MRF), which provide powerful and flexible spatial-contextual models for the prior distribution in Bayesian image analysis [53–55]. They have been recently used for HSI classification in conjunction with SVM [3, 56–58], active learning [59], multinomial logistic regression (MLR) [56, 60], subspace projections [57], hierarchical statis- tical region merging [61], blind source separation and mean- field approximations [62], multidimensional wavelets [63], sparse modeling and Dirichlet distributions [64], and ensemble classifiers [65, 66]. In [61] and [67], MRF-based methods were also developed for HSI segmentation. A further class of proba- bilistic graphical models is given by conditional random fields (CRF), which model as Markovian the posterior distribution directly [68]. HSI image classification methods have recently been developed using CRFs along with SVM and Mahalanobis distances [8, 69], MLR [70], decision tree ensembles [71], extreme learning machines [72], deep belief networks [73], segmentation and object-based image analysis [74], game theory [75], and adaptive differential evolution for decision fusion with LiDAR data [76]. Here, we shall focus on MRFs, first reviewing the basics and then discussing advanced meth- ods that integrate the MRF and SVM approaches to HSI classification.
Show more

32 Read more

PerTurbo manifold learning algorithm for weakly labelled hyperspectral image classification

PerTurbo manifold learning algorithm for weakly labelled hyperspectral image classification

In our context, another path of interest to improve the classification performances is to focus on generative algorithms, that naturally entail a description of each class. Of course, a cornerstone of the sta- tistical learning theory [5] is that learning a boundary between classes (as it is done with discriminative classifiers) is a simpler problem than training a model for each of them. As a direct consequence, one expects a classifier based on generative models to be less efficient than an optimal margin classifier such as SVM. However, from a theoretical point of view, generative learning is often more efficient than discriminative learning when the number of features is large compared to the number of training samples [19] while discriminative models are often better asymptotically. One reason is that the latter tends to overfit when the number of training examples is low (more practically, achieving an appropriate regularisation or tuning of SVM becomes difficult with few data). Ideal classifier is therefore likely to change when the number of training samples increases, and some attempts have been done to determine the threshold above which the classifiers should be switched in an on-line setting [20]. In the context of hyperspectral image classification, generative classifiers are particularly interesting since, as pointed out in [1], “supervised classification faces challenges related with the unbalance between high dimensionality and limited availability of training samples”. Nevertheless, comparison of the two models is a perennial topic as the superiority of generative classifiers may not be systematically observed, depending on the data and model specification [21]. Indeed, the first attempts in this direction in the remote sensing community were not successful so far, as illustrated in [22] where SVM still outperform kernel Fisher discriminant analysis when the training set size is low.
Show more

9 Read more

GWENN-SS: a simple semi-supervised nearest-neighbor density-based classification method with application to hyperspectral images

GWENN-SS: a simple semi-supervised nearest-neighbor density-based classification method with application to hyperspectral images

is found for a size of 31×31. The CVR index is also significantly better (lower) for GWENN-SS than for ss-Kmeans++ in all cases, which reveals a better ability of GWENN-SS to form coherent classes in the high dimensional space. However, the MSE index is lower for ss-Kmeans++; this result indicates that MSE, and to a more general extent the classification approaches which aim to optimize it (among which centroid-based methods) are not appropriate to HSI data due to the fact the class-conditional distributions are elongated and/or non-convex. In most cases, GWENN-SS is able to retrieve the correct labels of the actual GT, as well as to detect new classes which are consistent with it. Figure 3 shows some classification maps given by ss-Kmeans++ and GWENN-SS for various patch sizes. Notice that the maps shown for ss-Kmeans++ correspond to the result providing the best kappa index among the 10 runs of this method. Focusing on classes C1 and C2, where only C2 is present in the LS set, one can notice that for patch sizes 7×7 and 27×27, ss-Kmeans++ is able to discover a new class corresponding to C1, but with some confusion with C2; however, for a larger patch size (39×39), both classes are merged into C2. Comparatively, GWENN-SS provides much well-conditioned results, and can unveil the C1 class whatever the patch size, with very little confusion. As another example, class C3 is incorrectly merged with other classes according to the actual GT map by ss-Kmeans++; besides, even if GWENN-SS could not maintain the correct label for this class for the 7×7 patch, a consistent class was discovered in the corresponding region. Moreover, for larger patches, the label assigned was correctly maintained and propagated only to unlabeled pixels, without interference with other classes. Another important feature of GWENN-SS is its ability to produce classification maps which are spatially almost regular, in comparison with ss-Kmeans++. This property is all the more surprising as the classification algorithm does not account for spatial relationships between adjacent pixels. While it has not been proved yet, we conjecture that this is another consequence of the capability of a NN-DB classification method like GWENN to cope with elongated/non-convex distributions.
Show more

10 Read more

Synergetics Framework for Hyperspectral Image Classification

Synergetics Framework for Hyperspectral Image Classification

In this paper a new classification technique for hyperspectral data based on synergetics theory is presented. Synergetics – originally introduced by the physicist H. Haken – is an interdisciplinary theory to find general rules for pattern formation through self- organization and has been successfully applied in fields ranging from biology to ecology, chemistry, cosmology, and thermodynamics up to sociology. Although this theory describes general rules for pattern formation it was linked also to pattern recognition. Pattern recognition algorithms based on synergetics theory have been applied to images in the spatial domain with limited success in the past, given their dependence on the rotation, shifting, and scaling of the images. These drawbacks can be discarded if such methods are applied to data acquired by a hyperspectral sensor in the spectral domain, as each single spectrum, related to an image element in the hyperspectral scene, can be analysed independently. The classification scheme based on synergetics introduces also methods for spatial regularization to get rid of “salt and pepper” classification results and for iterative parameter tuning to optimize class weights. The paper reports an experiment on a benchmark data set frequently used for method comparisons. This data set consists of a hyperspectral scene acquired by the Airborne Visible Infrared Imaging Spectrometer AVIRIS sensor of the Jet Propulsion Laboratory acquired over the Salinas Valley in CA, USA, with 15 vegetation classes. The results are compared to state-of-the-art methodologies like Support Vector Machines (SVM), Spectral Information Divergence (SID), Neural Networks, Logistic Regression, Factor Graphs or Spectral Angle Mapper (SAM). The outcomes are promising and often outperform state-of-the-art classification methodologies.
Show more

6 Read more

A hyperspectral image classification algorithm based on atrous convolution

A hyperspectral image classification algorithm based on atrous convolution

In recent years, many researchers have used Convo- lutional Neural Networks (CNN) to classify hyper- spectral images and achieved good results. CNN learn the feature maps of samples through convolution and down-sampling hierarchical operations. Through mul- tiple feedback optimizations, it automatically learns and finally obtains hierarchical features. In particular, CNNs have developed towards the direction of “deep” and many classic architectures. For instance, AlexNet [10] VGG16, GoogleNet [11], and ResNet [12] can achieve good results in target recognition and classifi- cation on huge datasets. Usually, those datasets (e.g., ImageNet, PASCAL VOC) are composed of tens of thousands of samples, i.e., the spatial features of three dimension (RGB) images. These Deep Convolutional Neural Networks (DCNN) deepen with the increase of parameters and computation; thus, the training process becomes more difficult [13–15]. Therefore, how to construct a lightweight model has become a hot research topic. To achieve this goal, models such as MobileNets [16], Squeezenet [17], Xception [18], and others [19, 20] use deep-wise separable convolu- tion or dilated convolution instead of full convolution to make them lightweight and effective.
Show more

12 Read more

Show all 10000 documents...

Related subjects