CN-122023807-A - Aggregate contour automatic characterization method and equipment based on rapid semantic segmentation

CN122023807ACN 122023807 ACN122023807 ACN 122023807ACN-122023807-A

Abstract

The invention discloses an automatic aggregate contour characterization method and equipment based on quick semantic segmentation, and relates to the field of machine vision and building materials. The method is characterized in that the hardware system is stably collected and the software algorithm is efficiently processed. In terms of hardware, an image acquisition system integrated with a groove format bipartite, a high-definition high-speed camera array and a control computer is constructed, so that clear image sequences of aggregate are ensured to be obtained in the dynamic falling process. In terms of software, a kind of software is designed and implemented Is a deep learning segmentation algorithm. The algorithm is as follows Is a skeleton, and at least one structured state space based sequence model and at least one structured state space based sequence based model are embedded in the encoder and decoder paths of the encoder-decoder structure Constructed The module replaces the traditional And (5) a module. Through image acquisition, The method has the advantages of rapid semantic segmentation, automatic processing of morphological parameters and process integration, and automatic completion of morphological parameter extraction and report generation.

Inventors

SHI JIACHEN
SUN TUANWEI
ZHANG LI
CAI MEILING
XI CHENCHEN
SHI WEIBO
YAN SHOUJING
WEI JINTAO
Wang Shuailin
HU XUQUAN

Assignees

浙江省交通运输科学研究院
温州交发大桥有限公司

Dates

Publication Date: 20260512
Application Date: 20260211

Claims (10)

1. An automatic aggregate contour characterization method based on rapid semantic segmentation is characterized by comprising the following steps: The method comprises the steps of firstly, image acquisition, namely controlling a high-definition high-speed camera arranged in front of a groove format bipartite by a computer, and acquiring a digital image sequence in the process that aggregate freely falls down when passing through the groove format bipartite; step two, based on The rapid semantic segmentation of (1) is that the digital image of the aggregate collected in the step one is input into a pre-trained one In the deep learning semantic segmentation network, a binary segmentation map of an aggregate contour is output, wherein foreground pixels represent aggregates and background pixels represent non-aggregate areas; The network is based on Unet encoder-decoder architecture and merges in at least one of its encoder second, third, fourth downsampling stages and decoder first, second, third upsampling stages Module for replacing the tradition of the corresponding phase A module; Automatically processing morphological parameters, namely analyzing a connected region of the binary segmentation map obtained in the second step to identify each independent aggregate particle, and calculating one or more morphological parameters of each identified aggregate particle connected region, wherein the morphological parameters comprise projection area, long-short axis ratio, equivalent ellipse eccentricity and sphere diameter or sphere diameter based equivalent Particle size distribution of diameter; Step four, system integration and automation flow, namely, through one integrated type Application program, image acquisition and basis The rapid semantic segmentation and morphological parameter automatic processing of the system are controlled and managed in a flow manner, and full-automatic operation from aggregate image acquisition to parameter report generation is realized.
2. The automatic characterization method of aggregate contours based on rapid semantic segmentation according to claim 1, wherein: In the second step, the Mamba-Unet network uses the encoder-decoder structure of Unet as a base frame, and uses the encoder-decoder structure in at least one stage of the encoding path and decoding path The module replaces the original convolutional neural network module at the stage, and the convolutional neural network module The module is a visual feature extraction module constructed based on a structured state space sequence model and a self-attention mechanism and is used for improving the overall modeling efficiency and the segmentation speed of the model on the long-sequence dependency relationship of the image.
3. In the second step, the Mamba-Unet network is specifically constructed in such a way that the encoder comprises four downsampling stages, and one downsampling stage is respectively embedded after the second downsampling stage, the third downsampling stage and the fourth downsampling stage The decoder comprises four upsampling stages, one embedded after the first upsampling stage, the second upsampling stage and the third upsampling stage And the encoder and the decoder are connected in a jump mode to perform feature fusion.
4. The automatic characterization method of aggregate contours based on rapid semantic segmentation according to claim 1 or 2, characterized in that: In the second step, the The network composition is as follows: (a) The said The module is composed of two parts, and the processing sequence according to the data is as follows: A second part composed of a self-attention mechanism and a multi-layer perceptron; (b) MambaVision Mixer is a visual feature extraction module constructed based on a structured state space sequence model, the structured state space sequence model core passing through a state equation And output equation The present feature map, wherein x (t) is an input feature sequence, h (t) is a hidden state sequence, y (t) is an output feature sequence, A, B, C is an optimized coefficient matrix in the network training process; (c) Self-attention mechanism passes through vector sets and equations thereof And Feature mapping is implemented, where x is the input feature sequence, y is the output feature sequence, As a matrix of coefficients that can be trained, Query vectors, key vectors, and value vectors, respectively, softmax is a normalized exponential function.
5. The automatic characterization method of aggregate contours based on rapid semantic segmentation according to claim 1, wherein: in the second step, the training process of the Mamba-Unet network includes: (a) The data set construction comprises the steps of collecting and labeling to form a plurality of aggregate sample images and corresponding pixel-level binary labeling diagrams thereof, and forming a training set, a verification set and a test set; (b) Data enhancement, namely, in the training process, one or more operations of random rotation, random clipping, brightness contrast adjustment and adding telling noise are synchronously applied to an input image and a label thereof so as to enhance the generalization capability of a model; (c) Model training, namely building a network based on Pytorch frames, setting an Adam optimizer, setting an initial learning rate eta, and performing multiple rounds of iterative training on a computing platform by using a composite loss function until the performance index of the model on a verification set is converged.
6. The automatic characterization method of aggregate contours based on rapid semantic segmentation according to claim 1, wherein: In the third step, morphological parameters are identified: (a) The projection area is equivalently represented by the number of pixel points corresponding to the segmented aggregate area; (b) By carrying out ellipse fitting on the aggregate particle connected region, taking the ratio of the length of the major axis to the length of the minor axis of the fitted ellipse as the ratio of the major axis to the minor axis, and based on the geometric relationship between the focal point and the major axis of the ellipse Calculating an equivalent elliptic eccentricity e, wherein a is a major axis half length and b is a minor axis half length; (c) Calculating particle size distribution in the third step, and outputting a particle size distribution curve in a histogram or cumulative distribution function form by counting the equivalent circle diameters of all particles, wherein the equivalent circle diameters are , Is the projected area.
7. An automatic aggregate contour characterization device based on rapid semantic segmentation for implementing the automatic aggregate contour characterization method based on rapid semantic segmentation as set forth in any one of claims 1 to 5, comprising: (a) The groove format bipartite is used for receiving and guiding aggregate to be tested to form uniform and dispersed lower blanking flow; (b) The image acquisition subsystem comprises a high-definition high-speed camera which is arranged right in front of the trough-format bipartite and is used for capturing images of falling aggregates; (c) A computing processing subsystem comprising at least one computing device configured with a processor and a memory storing a computer program that when executed by the processor implements functions comprising the Mamba-Unet deep-learning semantic segmentation network, a connected region analysis algorithm, and the Windows application.
8. The automatic aggregate contour characterization device based on rapid semantic segmentation according to claim 6, further comprising a solid background plate for placing a lower part of the trough divider, wherein the solid background plate provides a uniform and stable background for the digital image to facilitate aggregate segmentation.
9. The apparatus of claim 6, wherein the high-definition high-speed camera has a resolution of not less than 1920x1080 pixels, and a frame rate of not less than 200 frames/sec using a global shutter to accommodate clear capture of aggregate high-speed drops.
10. The apparatus of claim 6, wherein the Windows application in the computing subsystem provides a graphical user interface comprising at least a real-time image display window, an acquisition control button, a segmentation result visualization area, a morphological parameter data table, and a particle size distribution chart display area, and data export function options.

Description

Aggregate contour automatic characterization method and equipment based on rapid semantic segmentation Technical Field The invention relates to application of computer vision in the field of building materials, in particular to an image semantic segmentation technology, and specifically relates to a high-efficiency semantic segmentation algorithm combined with a novel state space sequence model, and a method and complete equipment for automatically and highly precisely characterizing aggregate particle needle sheets and other contours in the building materials by utilizing the algorithm. Background Aggregate is used as a main component of composite materials such as concrete, asphalt mixtures and the like, and physical morphological parameters such as the size (particle size distribution), shape (angular, sphericity, flatness), surface texture and the like of particles are key factors for determining the mechanical properties, workability and durability of the final composite materials. For example, the size distribution of the aggregate directly affects the mix design and compactness of the concrete, and the shape characteristics (such as long-short axial ratio and eccentricity) of the particles are closely related to the compressive strength of the aggregate, interlocking between the particles and the fluidity of the mixture. At present, the characterization method of aggregate outline morphology is mainly divided into two main categories, namely traditional manual measurement and automatic measurement based on image processing. The traditional manual measurement generally depends on mechanical methods such as vernier calipers and screening, so that the time and the labor are consumed, the efficiency is extremely low, larger human errors and subjectivity exist, shape parameters are difficult to obtain, and the requirement of large-scale quality detection cannot be met. With the development of computer technology, an automatic measurement method based on image processing is gradually rising. Such methods typically take an image of the aggregate by an industrial camera and then use digital image processing algorithms (e.g., edge detection, thresholding, morphological operations, etc.) to extract the particle contours and calculate morphological parameters. However, such conventional image processing methods still mainly collect images after manually adjusting the position of the aggregate, and have low efficiency. And aggregate images often present challenges of (1) uneven illumination leading to shadows and reflections, and (2) low contrast of particle color, texture, and background. The traditional algorithm needs to carefully adjust parameters for specific scenes, has poor generalization capability, is particularly non-ideal for the segmentation effect of the adhesion particles, and easily causes serious deviation of counting and size measurement. In recent years, deep learning techniques, particularly semantic segmentation networks such as FCN, U-Net, deepLab, have been revolutionized in the field of image segmentation. U-Net, by virtue of its symmetrical encoder-decoder structure and jump connection, is excellent in biomedical image segmentation, and has also been tried for segmentation of rock, aggregate. However, the direct application of U-Net to aggregate segmentation remains a challenge. The encoder of U-Net is usually based on Convolutional Neural Network (CNN), its inherent local connection characteristics have limitations in modeling image global context information, and for aggregates with extremely irregular shape and complex contour, it is difficult to achieve the most accurate segmentation. On the other hand, although the recently emerging transducer architecture has global modeling capability through a self-attention mechanism, the computational complexity is proportional to the square of the image size, and the computational overhead is huge when processing high-resolution aggregate images, so that the requirements of fast and real-time processing on an industrial site are difficult to meet. In terms of hardware systems, most of existing aggregate image acquisition devices are static shooting or simple dynamic shooting, manual arrangement is often required, time and labor are wasted, the degree of automation is low, stable and clear images are difficult to acquire when aggregates fall at a high speed, and the accuracy of subsequent analysis is affected. Disclosure of Invention In order to achieve the aim, the invention adopts the following technical scheme that the automatic characterization method of the aggregate profile of the rapid semantic segmentation is characterized by comprising the following steps: The method comprises the steps of firstly, image acquisition, namely controlling a high-definition high-speed camera arranged in front of a groove format bipartite by a computer, and acquiring a digital image sequence in the process that aggregate freely falls down when passing throug