On-line Visual Recognition Of Coal Gangue Based On BLOB Analysis And Machine Learning

Coal gangue sorting is an important means to ensure the quality of coal.At present,the gangue sorting is mainly manually operated by worker in the coal mine factory.It is not only inefficient,but also brings strong labor intensity.Bad working environment could also do harm to the health of workers.Therefore,it is urgent to develop an intelligent coal gangue sorting system to replace manual operation.At present,X-ray method and image processing method based on machine vision are main recognition technologies in an intelligent coal gangue sorting system.X-ray recognition is only suitable for small gangue sorting because it is difficult to penetrate large gangue block.Moreover,the radioactivity of radiation,would do harm to the health of workers.At present,many image recognition technologies based on machine vision are mainly for static image analysis,not considering about the bad influence of conveyor belt and the moving speed.Black conveyor belt and the shape of gangue on the actual pipeline have a great impact on gangue imaging,resulting in a smooth gray value transition especially near the gangue edge.Because of this,it's very difficult to position the object edge by usual image processing methods.Coal gangue recognition and positioning become the key problem in the intelligent gangue sorting system.Firstly,regarding the medium and large gangue block as a specific detection target,a machine vision system for coal gangue recognition is designed.The components of vision system are selected and designed by focusing on the analysis of detection target and recognition requirements.According to the size of the target gangue,the belt width of the conveyor,the transmission speed and the installation height of the vision system,the camera and lens are selected to ensure a suitable vision field?quick image acquiring under the moving state and the sufficient resolution.Coal gangues have various size?different shapes and similar colors.Regarding all of these,a large area surface light source controlled by constant current source is designed.The light source is arrayed with high-power LED.The electrical connection between LEDs is in series and parallel.The outermost layer of the light source is arranged with milky white acrylic plate,which plays the role of light diffusion and can ensure that the light source has large light diffusion angle.These designs ensure that the light source not only has enough brightness and evenness,but also has good adaptability.Using this light source in a vision system,the coal gangue or coal with different size,height and shape will not be affected by the light to form a high brightness reflection area and edge contour shadow,which ensuring the system to obtain a clear image.In order to acuquire clear image with high contrast,vision system for gangue recongnition needs manual or automatic focusing before actual use.Image definitionevaluation function is mainly used to evaluate whether the vision system is in the focus position.Firstly,the basic theoretical basis of each clarity evaluation function is analyzed from the perspective of optical imaging.The image definition evaluation functions are studied,focusing on Fourier evaluation function?DCT evaluation function and their improved weighted evaluation function.Focusing characteristics are compared beteween the definition evaluation function and its weighting function based on frequency domain transformation.The tests are carried out by using the circuit board and calibration template as focusing objects.The influence of image content,light intensity,filter radius,focus step distance and focus window on the frequency domain evaluation function is studied.The test results show that the improved DCT evaluation function has good unbiased,monotonic and unimodal performance,good sensitivity near the focus,good evaluation effect for different evaluation windows and good algorithm stability.The improved definition evaluation function is used to assist manual focusing on the gangue sorting system,which reduces the difficulty of manual focusing and the number of adjustments.In order to locate coal gangue and get its geometric parameters,a BLOB feature extraction algorithm based on image threshold segmentation and run length connected labeling is developed by analyzing and researching the existing algorithms.Firstly,the segmentation threshold is calculated,and then it's used to segment the image.The BLOB partitions are divided by run length connected labelling method.A lot of feature parameter values,including the distribution range,central coordinates and centroid coordinates,are gotten by using the information of all the pixels in the same blob.The algorithm has the advantages of small storage,low complexity,fast computation and high search performance.The geometric feature information including the distribution range of coal gangue could be extracted easily by this method,which solved the problem of gangue edge location.In order to solving the problem that BLOB analysis algorithm takes too much computing time and memory capacity when analyzing an image with a lot of blobs,a gangue recognition method based on non-gangue image filtering and BLOB analysis is developed.The third-order moment is used as filtering evaluation parameter.Most of the images without gangue are filtered by setting filtering threshold after image evaluation.This method greatly reduces the invalid computing time.For the suspected gangue image,BLOB analysis method is used to locate the target position and extract geometric features.Through locating the target distribution range,the blob area size and the third-order moment evaluation of local area image content are used to make an accurate judgment,which reduces the adverse impact of background on judgment and improves the recognition accuracy.On this basis,the software of intelligent gangue sorting system is designed.The research algorithm is applied to the large-scale experimental equipment for gangue sorting.The real-time identification and location of coal gangue are realized.The equipment runs steadily and reliably.The gangue sorting experiment get good result and gangue sorting rate reaches 91.1%.In order to further improve the recognition accuracy of low probability gangue,the experiment of low probability gangue recognition based on machine learning was carried out.ROI images are obtained by BLOB analysis to low probability gangue and coal image,which are used as the traing samples.The sample set is constructed by expansion method.Nine evaluation parameters are selected to evaluate the features of sample set images and data sets are built.K-CV method is used to optimize the parameters of support vector machine model.By using support vector machine classification method,the correct classification model is trained and established.After that,the test samples are predicted,and the prediction accuracy of small probability gangue samples is 100%.The simulation results show that the method is effective in improving the recognition rate of low probability gangue.