| Coal,feeding the industrial development,is always the main energy of China.In recent years,China has been making considerable efforts to reform the coal supply front and advocate the development of clean energy.As a result,the proportion of coal consumption to total energy consumption is constantly decreasing.In China,however,coal,oil,and gas is rich,deficient,and lean,respectively.These resource endowment features can hardly change the trend that coal continues to be dominent among all kinds of energy sources even in a very long future in China.Automatical coal-rock recognition remains to be an open world-class problem,which is one of the key technologies for intelligent coal mining and processing.Addressing it will benefit the adaptive height adjustment of the shearer’s drum,the process control of fully mechanized top-coal caving,and the fast coal-gangue separation in coal preparation plants.Most traditional coal-rock recognition methods have some drawbacks,such as unsatisfactory performance,lack of robustness,narrow scope of application.In view of the fact that coal and rock have distinctive visual features,this work aims to explore some novel approaches to coal-rock recognition via visual computing.It was jointly supported by the National Natural Science Foundation of China(Grant NO.51134024),the National Key Research and Development Program of China(Grant NO.2016YFC0801800),and the National High Technology Research and Development Program(i.e.National 863 Program)of China(Grant NO.2012AA062203).The main contributions and innovations of this work are four-fold,and they are drawn as follows.1)A coal-rock recognition method based on wavelet-domain asymmetric generalized Gaussian models is proposed.Firstly,multi-level discrete wavelet transform is conducted on coal and rock images.As a consequence,several medium and high frequency subbands can be obtained.Then,statistical modelling of the subband coefficients using asymmetric generalized Gaussian distribution is done.Maximumlikelihood estimator is employed to estimate the parameters of asymmetric generalized Gaussian distribution.Finally,having relative entropy with respect to asymmetric generalized Gaussian distribution been formulated,coal-rock recognition is completed using relative entropy based similarity measurement.The results of several random sampling experiments demonstrate that the recognition accuracy rate the proposed method achieves is more satisfactory than those other similar ones do.Accordingly,the proposed method is further modified.The modified method adopts steerable pyramidal decompostion instead of discrete wavelet transform since the former is much more direction selectable and rotation invariant compared with the latter.Experimental results show that the modified method leads to some improvement over the originally proposed one(i.e.the unmodified one)in terms of recognition accuracy rate.It is reasonable to suppose that the subband coefficients of coal and rock images generated by discrete wavelet transform or steerable pyramidal decomposition obey asymmetric generalized Gaussian distribution.Additionally,experimental results also reveal that both the originally proposed method and the later modified one satisfy the real-time requirement.2)A coal-rock recognition method using dual-tree complex wavelet transform and generalized gamma distribution is proposed.Firstly,coal and rock images are decomposed using dual-tree complex wavelet transform,which leads to several directional subbands.Next,a simple yet effective strategy for rotation invariance enhancement is proposed to address the problem of similarity matching of subband misplacement caused by sample rotation.In accordance with this strategy,six directional subbands generated after every level dual-tree complex wavelet transform are sorted in descending order by the product of mean and variance of absolute coefficient.Then,statistical modelling of the absolute coefficients of directional subbands based on generalized gamma distribution is done.The parameters of statistical model are estimated based on scale-independent-shape-estimation equation.Finally,having relative entropy with respect to generalized gamma distribution been formulated,automatical identification of coal and rock images is completed using relative entropy based similarity measurement.The results of several both cross validation experiments and random sampling experiments demonstrate that the classification performance of the proposed approach is more satisfactory compared with other similar methods.The strategy for rotation invariance enhancement is beneficial to the proposed method.On one hand,its recognition accuracy rate can be further improved by means of this strategy.On the other hand,based on this strategy,making a trade-off among recognition accuracy rate,computational time,and storage requirement is convenient for engineers or technicians.In addition,experimental results confirm that statistical modelling of the absolute coefficients of directional subbands of coal and rock images generated by dual-tree complex wavelet transform with generalized gamma distribution is practicable.3)The state-of-the-art median filtering based robust extended local binary patterns(RELBP)is introduced into the coal-rock recognition community.The importance distribution of the classification-oriented RELBP features of coal and rock images is of imbalance.(i.e.The weight matrix with respect to the importance of the classificationoriented RELBP features of coal and rock images is structurally sparse.)Having the optimization problems involved in least squares regression model and regularized correntropy model been solved,an effective approach to coal-rock recognition via optimal selection of RELBP using least squares regression,and another effective coal-rock recognition method based on RELBP and regularized correntropy model are presented,respectively.Firstly,the RELBP-based native feature vectors are extracted from coal and rock images.Next,a least squares regression model or a regularized correntropy model is established on the native feature vectors of training samples.By solving the problem corresponding to the least squares regression model or the regularized correntropy model,optimal patterns almost as discriminative as native patterns are obtained.Then,based on these optimal patterns,the RELBP-based optimal feature vectors are extracted from their corresponding native ones.Finally,the completion of coal-rock recognition depends on the chi-square distance between two optimal feature vectors of test sample and training one.The results of several cross validation experiments show that,the presented two methods not only have significant improvements over most of other approaches in terms of recognition accuracy rate,but also yield slightly higher recognition accuracy rates than the off-the-shelf method directly using RELBP-based native feature vectors(i.e.the RELBP method).The results of several random sampling experiments show that,in terms of recognition accuracy rate,these two methods are comparable to the RELBP method,whereas they substantially outperform other methods.It is worth mentioning that,the storage requirement for these two methods is a bit looser than that of the RELBP method.Although solving the optimization problem that corresponds to the least squares regression model or the regularized entropy model is time-consuming,it is completed off-line at the sample training stage.No matter how long the duration of the optimization problem solving is,it is always available to make these two methods real-time.4)As an image feature descriptor,RELBP is novel enough,but its parameter settings are a little bit cumbersome and the computation is relatively time-demanding.Hence,research interests move to another descriptor,completed local binary patterns(CLBP),the discriminative power of which is acceptable and the computation of which is much faster compared with RELBP.According to the suggestion concerning the parameter settings of CLBP,the dimension of CLBP-based native feature vector is much higher than that of RELBP-based native one.To deal with this problem,some transformation is performed on CLBP-based native feature vectors of coal and rock images,an effective approach to coal-rock recognition via support vector guided dictionary learning of CLBP,and another coal-rock recognition method based on the class-specific reconstruction residual of CLBP are proposed.The basic workflow of the former method is summarized as follows.Firstly,the multi-scale CLBP-based feature vectors of coal and rock images are extracted.Next,support vector guided dictionary learning is conducted on the CLBP-based feature vectors of category-known training samples.One dictionary for coal-rock characterization,several weight vectors and biases are obtained after that.Support vector guided dictionary learning not only lets the dictionary-based representations(i.e.coding vectors)of CLBP-based feature vectors of training samples be sparse to some extent,but also makes these representations be as linearly separable as possible.Then,the dictionary-based coding vector of CLBP-based feature vector of test sample is gained.Finally,a linear discriminant function is employed to complete the classification of the coding vector of test sample.The weight vectors and biases are two kinds of parameters of this function.The basic workflow of the latter method is summarized as follows.Firstly,the multi-scale CLBP-based feature vectors of coal and rock images are extracted.Then,projective dictionary pair learning is conducted on the CLBP-based feature vectors of category-known training samples.Several pairs of class-specific dictionaries are acquired after that.One class corresponds to one pair of dictionaries(i.e.synthesis dictionary,and analysis dictionary).Finally,different dictionary pairs,corresponding to different classes,are used to reconstruct the CLBP-based feature vector of test sample.The category corresponding to the dictionary pair that yields minimum reconstruction residual is treated as the result of coal-rock recognition.The results of several both cross validation experiments and random sampling experiments reveal that,these two methods work better than others.Especially,they can achieve surprisingly high recognition accuracy rates even at training sample-insufficient random sampling experiments.Although both support vector guided dictionary learning and projective dictionary pair learning involved in these two methods are time-consuming,they are completed at the sample training stage.Therefore,only CLBP-based feature extraction and category discrimination are essential factors affecting the real-timeness of these two methods.Experimental results show that these two methods run in real time at both cross validation experiments and random sampling experiments.Additionally,the storage requirements for them are independent of the number of training samples.To some extent,this facilitates the hardware-system integration of them in the future.Those hardware systems integrating with the proposed two methods can adapt themselves well to the circumstances that the number of training samples varies over time. |
PDF Full Text Request