Bulk material particle size on-line detection method and system

Abstract

The application discloses a kind of bulk material particle size on-line detection method and system, belong to bulk material particle detection technical field.The method includes: to the bulk material particle flow to be detected is continuously imaged at high speed, obtains particle sequence image;Particle sequence image is carried out based on edge keeping filtering gray processing, obtains particle gray image;Based on watershed algorithm, particle contour segmentation is carried out to particle gray image, obtains initial particle mask set;Initial particle mask set is input into the discriminator of the deep convolution generative adversarial network that is constructed in advance, the authenticity of each particle mask is discriminated by discriminator, and the particle mask higher than preset discrimination threshold is screened out, and constitute refined particle mask set;Equivalent ellipse fitting is carried out to each particle mask in refined particle mask set, and the equivalent long axis length and equivalent short axis length of each particle are obtained.

Claims

1. A method for online detection of particle size of bulk materials, characterized in that, The method includes: High-speed continuous imaging of the particle flow of the bulk material to be tested is performed to obtain particle sequence images; The particle sequence image is subjected to grayscale processing based on edge-preserving filtering to obtain a particle grayscale image; The particle grayscale image is segmented based on the watershed algorithm to obtain an initial set of particle masks. The initial set of particle masks is input into the discriminator of a pre-constructed deep convolutional generative adversarial network. The discriminator performs a realism judgment on each particle mask in the initial set of particle masks, and filters out particle masks that are higher than a preset judgment threshold to form a refined set of particle masks, specifically including: A training set of particle masks labeled with real particle contours is pre-collected. The training set of particle masks is used to perform adversarial training on a deep convolutional generative adversarial network. The generator of the deep convolutional generative adversarial network is used to generate simulated particle masks, and the discriminator is used to determine whether the input mask is a real particle mask. Each particle mask in the initial set of particle masks is scaled to a fixed size required by the input layer of the discriminator of the deep convolutional generative adversarial network to obtain a standardized particle mask. The standardized particle mask is input into the discriminator after training, and the discriminator outputs a authenticity score between zero and one. The original size particle mask corresponding to the standardized particle mask with a realism score greater than 0.8 is retained, and all the retained original size particle masks constitute a refined particle mask set. For each particle mask in the refined particle mask set, perform equivalent ellipse fitting to obtain the equivalent major axis length and equivalent minor axis length of each particle, specifically including: Calculate the zeroth and first moments of each particle mask in the set of refined particle masks to obtain the centroid coordinates of the particle mask; Calculate the second central moment of the particle mask, construct the covariance matrix based on the second central moment, and perform eigenvalue decomposition on the covariance matrix to obtain the first eigenvalue and the second eigenvalue. Four times the square root of the first eigenvalue is taken as the equivalent major axis length, and four times the square root of the second eigenvalue is taken as the equivalent minor axis length. A particle size distribution matrix is ​​constructed based on the equivalent major axis length and equivalent minor axis length of all particles. This particle size distribution matrix is ​​then output as the online detection result of the bulk material particles, specifically including: Search for the minimum and maximum equivalent major axis lengths from the equivalent major axis lengths of all particle masks, and search for the minimum and maximum equivalent minor axis lengths from the equivalent minor axis lengths of all particle masks. Using the equivalent major axis length as the basis for dividing the horizontal axis interval, the range from the minimum equivalent major axis length to the maximum equivalent major axis length is evenly divided into a preset number of granular intervals; Using the equivalent minor axis length as the basis for dividing the vertical axis interval, the range from the minimum equivalent minor axis length to the maximum equivalent minor axis length is evenly divided into the same preset number of minor axis intervals; The number of particles corresponding to each particle size interval and each minor axis interval is counted, and a particle size distribution matrix with a preset number of rows and columns is constructed. The element in the i-th row and j-th column of the particle size distribution matrix represents the number of particles whose equivalent major axis length falls in the i-th particle size interval and whose equivalent minor axis length falls in the j-th minor axis interval.

2. The method according to claim 1, wherein, The process of performing high-speed continuous imaging of the particle flow of the bulk material to be detected to obtain particle sequence images specifically includes: A high-speed industrial camera continuously acquires images of a moving bulk material particle stream at a preset sampling frame rate. The main optical axis of the high-speed industrial camera is perpendicular to the direction of movement of the bulk material particle stream. During the acquisition process, a strobe light source is synchronously triggered to provide supplementary lighting, and the flash frequency of the strobe light source is the same as the sampling frame rate of the high-speed industrial camera. The acquired continuous images are stored as particle sequence images in time stamp order, and each frame of the particle sequence image contains at least one complete bulk material particle.

3. The method according to claim 2, wherein, The particle sequence image is subjected to grayscale processing based on edge-preserving filtering to obtain a particle grayscale image, specifically including: The particle sequence image is converted from the RGB color space to the HSV color space, and the luminance component in the HSV color space is extracted as the initial luminance image. An edge-preserving filter is applied to the initial brightness image using a guided filtering algorithm. The guided image for the guided filtering algorithm is the initial brightness image itself, and the filtering radius parameter is set to three times the number of pixels corresponding to the smallest particle diameter in the particle sequence image. The brightness image after guided filtering is output as a grainy grayscale image.

4. The method according to claim 3, wherein, The step of segmenting the particle grayscale image using the watershed algorithm to obtain an initial set of particle masks specifically includes: A morphological reconstruction operation is performed on the grainy grayscale image. The morphological reconstruction operation uses circular structuring elements for both open and closed reconstruction to obtain a foreground marker image after background suppression. Calculate the distance transformation for the foreground marker image to obtain a distance transformation map, wherein the value of each pixel in the distance transformation map represents the Euclidean distance from that pixel to the nearest background pixel; The distance transformation graph is inverted to obtain an inverted distance graph, and the local maxima of the inverted distance graph are used as seed points for the watershed algorithm. Using the seed point as the starting point, perform a watershed transformation on the inverted distance map to obtain the closed contour of each particle. Use the pixel region inside the closed contour as a single particle mask, and all particle masks constitute the initial particle mask set.

5. The method for online particle size detection of bulk materials according to claim 4, characterized in that, In the morphological reconstruction operation, both open and closed reconstruction use diamond-shaped structural elements with a radius equal to the number of pixels corresponding to the smallest particle radius in the particle sequence image.

6. The method for online particle size detection of bulk materials according to claim 5, characterized in that, A spectral normalization layer is added to the discriminator of the deep convolutional generative adversarial network. The spectral normalization layer normalizes the spectral norm of the weight matrix of each convolutional kernel in the discriminator.

7. An online particle size detection system for bulk materials, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the online particle size detection method for bulk materials as described in any one of claims 1 to 6.

Images