Particle Morphology Recognition Method Based on YOLOv7 and Interference Imaging Technology

Abstract

This invention discloses a particle morphology recognition method based on YOLOv7 and interferometric imaging technology. The method uses an interferometric particle imaging system to obtain interferometric defocused images of a mixed particle field to create a dataset. The annotated interferometric defocused images are divided into training, validation, and test sets. The network is then improved by incorporating a collaborative attention module to enhance detection accuracy. Hyperparameters are adjusted, and training and testing are performed using the annotated dataset. After 200 epochs of training, the optimal weights are obtained, ultimately achieving mixed particle field morphology recognition. This method can achieve high-precision identification of particle morphology in mixed particle fields, providing technical support for acquiring cloud particle field information.

Description

Technical Field

[0001] The present invention belongs to the field of image processing, and in particular relates to a particle morphology recognition method based on YOLOv7 and interference imaging technology. Background Art

[0002] Clouds and precipitation are crucial links in the water cycle and play a vital role in the Earth's water cycle. Liquid droplets and ice crystals in clouds can transform into each other. Based on the microstructural characteristics of cloud particles, clouds are classified as ice clouds, water clouds, and a mixture of the two. Acquiring information about the morphology and size of cloud particles is key to cloud microphysics measurements. Determining the morphology of cloud particles plays a key role in obtaining information about the water or ice content in clouds, which helps us more accurately predict the weather and avoid meteorological disasters. It is also of great help in revealing the formation and development of clouds and artificial rainfall. Currently, the most commonly used method for acquiring particle information is interferometric particle imaging (IPI).

[0003] Interferometric particle imaging was developed in 1986 et al. first proposed a measurement method based on the distribution of scattered light from particles. This method uses Mie scattering theory to study the relationship between particle diameter and its scattered light distribution. A light intensity detector is used to receive the scattered light from the particles, thereby obtaining particle size information. IPI technology offers advantages such as a wide measurement range, high accuracy, and non-contact operation, making it suitable for cloud microphysics measurements. In recent years, interferometric particle imaging (IPI) has been widely used in particle size and shape measurement, morphology recognition, and other applications. Patent CN105866013A proposes a method for distinguishing spherical particles using interferometric particle imaging, titled "Spherical Particle Discrimination Method Based on Two Laser Interference Imaging Defocus Interference Patterns." This method studies the polarization properties of transparent spherical and transparent ellipsoidal particles, proposing that when polarized light illuminates the particles, spherical particles retain their polarization, while the polarization state of polarized light changes after passing through ellipsoidal particles. A dual-receiving system was established: one receiving system directly images the particle, while a polarizer is added before the other receiving system to record the particle field interference defocus information. After the polarizer is added, only the defocused interference image of the ellipsoidal particle is visible on CCD2. The morphology of spherical and ellipsoidal particles can be distinguished by interfering defocused images under different polarization states. However, this method requires the use of a polarization device and is relatively cumbersome, making it difficult to obtain real-time cloud particle field information.

[0004] More and more disciplines are beginning to integrate with deep learning. The YOLO (You Only Look Once) series is a classic one-stage object detection network in deep learning. Compared to some two-stage networks (such as the R-CNN series), it offers higher performance and lower detection latency. YOLOv7, proposed by Chien-Yao Wang and Alexey Bochkovskiy in 2022, boasts several key innovations over YOLOv5: First, it employs multiple "free" training techniques during training, improving detection performance. "Free" training techniques are used only during training and do not increase inference time. Second, it features a more efficient feature extraction network, ELAN. Third, it utilizes model reparameterization and label assignment strategies. Fourth, it utilizes model scaling. However, currently, there are no existing examples of using YOLOv7 for particle morphology discrimination. The challenges of using this method include: first, the difficulty in obtaining datasets. Achieving morphological recognition in mixed particle fields requires the generation of extensive experimental data. Because spherical and rough particles must be within the same image plane, the training process requires diluting the spherical particles and discretely placing the rough particles on a glass slide to obtain defocused images. Second, network parameters and structure. Network hyperparameters such as the learning rate, cosine annealing hyperparameters, and learning rate momentum need to be adjusted. Next, network structure should be modified to enhance generalization and detection performance. Third, particle overlap. When there are a large number of particles, overlap occurs, which can negatively impact morphological recognition results. Summary of the Invention

[0005] The purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a particle morphology recognition method based on YOLOv7 and interferometric imaging technology to achieve high-precision particle morphology recognition function and provide technical guarantee for the measurement of cloud particle field.

[0006] The technical solution adopted by the present invention is: a particle morphology recognition method based on YOLOv7 and interferometric imaging technology, which is carried out in the following steps:

[0007] Step 1: Build an interference particle imaging system to obtain mixed particle interference defocused images. The interference particle imaging system includes: a laser, a spatial filter composed of an objective lens and a pinhole, a collimating lens, a lens group composed of a plano-convex cylindrical lens and a plano-concave cylindrical lens, a glass slide, an imaging lens, and a CCD camera; from left to right in the horizontal direction, a 532nm laser 1, a spatial filter composed of an objective lens 2 and a pinhole 3, a collimating lens 4, a lens group composed of a plano-convex cylindrical lens 5 and a plano-concave cylindrical lens 6, a glass slide containing polystyrene spherical particles and sand are arranged in sequence, and an imaging lens 8 and a CCD camera 9 are arranged below in sequence; use this system to obtain a mixed particle field interference defocused image dataset; place sand and diluted polystyrene spherical particles on a carrier, and obtain a mixed particle field interference defocused image dataset at the CCD end after being irradiated by a sheet laser beam;

[0008] Step 2: Label and divide the dataset. Use LabelImg software to label the mixed particle field interference defocused image dataset obtained in step 1. The annotation content includes particle type and particle position information, and the label format is YOLO (txt file format). After labeling, the mixed particle field interference defocused image dataset is divided into training set, validation set, and test set in a ratio of 7:1:2;

[0009] Step 3: Network Structure Improvement: First, three Coordinate Attention (CA) modules were added to the YOLOv7 network. Specifically, they were located in the YOLOv7 network head, after the ELAN feature extraction network at layers 75, 88, and 101. Second, the feature fusion module in the original YOLOv7 network was replaced with the Spatial Pyramid Pooling–Fast (SPPF) module, reducing the number of parameters and computational complexity and accelerating particle morphology detection.

[0010] Step 4: Network training to obtain optimal weights. Use the dataset annotated in Step 2 and the improved network from Step 3 for training. Hyperparameter settings are as follows: initial learning rate to 0.009, learning rate momentum to 0.935, weight decay coefficient to 0.0004, cosine annealing parameter to 0.1, bounding box loss weight coefficient to 0.05, anchor scaling threshold to 4.0, and IoU training threshold to 0.2. The training set divided in Step 2 is fed into the network for 200 epochs to obtain optimal weights.

[0011] The laser 1 emits parallel polarized light with a wavelength of 532 nm and a maximum output power of 3 W.

[0012] A CCD camera with a resolution of 2448×2048 and a pixel size of 3.45 μm is used as the receiving device.

[0013] The particles to be tested are 45 μm polystyrene spherical particles and sand particles, which are placed on a glass slide.

[0014] The present invention has the following technical effects:

[0015] First, real-time processing. By inputting a mixed particle field interference defocused image, the target's position and category can be directly calibrated. Compared with traditional image processing methods, this method has higher real-time performance.

[0016] Second, high precision: Overlapping affects the accuracy of morphological recognition, and this technology can overcome the impact of overlap on morphological recognition to a certain extent.

[0017] Third, it has strong generalization capabilities. Changes in the size of the interference fringe circles, the frequency of the fringe patterns, and even the size of the sand grains have no impact on the accuracy of morphological recognition. The model has strong generalization capabilities and can be applied to real-world scenarios.

[0018] The present invention uses YOLOv7 and interferometric imaging technology to perform particle morphology recognition, providing technical support for the acquisition of cloud particle field information. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the particle morphology discrimination method of the present invention;

[0020] Figure 2 Schematic diagram of the interferometric particle imaging system;

[0021] Figure 3 is the particle morphology recognition result. DETAILED DESCRIPTION

[0022] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0023] The particle morphology recognition method based on YOLOv7 and interference imaging technology of the present invention is as follows: Figure 1 As shown, specifically:

[0024] Step 1: Build Figure 2The interferometric particle imaging system shown here acquires interferometric defocused images of mixed particles. The interferometric particle imaging system includes: a laser, a spatial filter consisting of an objective lens and a pinhole, a collimating lens, a lens assembly consisting of a plano-convex cylindrical lens and a plano-concave cylindrical lens, an imaging lens, and a CCD camera. Horizontally, from left to right, the following are arranged: a 532 nm laser 1, a spatial filter consisting of an objective lens 2 and a pinhole 3, a collimating lens 4, a lens assembly consisting of a plano-convex cylindrical lens 5 and a plano-concave cylindrical lens 6, and a glass slide 7 containing polystyrene spherical particles and sand. Below glass slide 7, an imaging lens 8 and a CCD camera 9 are arranged in this order. This system acquires interferometric defocused image datasets of mixed particle fields. System parameters are set as follows: object distance is 190 mm, image distance is 80.21 mm, scattering angle is 90°, and magnification is 0.416. The particles to be measured are 45 μm polystyrene spherical particles and sand particles of varying sizes. Polystyrene spherical particles are diluted with alcohol, and sand particles are discretely placed on a glass slide using tweezers. After being irradiated by a sheet laser beam, a defocused image of the mixed particle field interference is obtained on the CCD end.

[0025] Step 2: Label and divide the dataset. Use LabelImg software to label the mixed particle field interference defocused image dataset obtained in step 1. The annotation content includes particle type and particle position information, and the label format is YOLO (txt file). After labeling, the mixed particle field interference defocused image dataset is divided into training set, validation set, and test set in a ratio of 7:1:2;

[0026] Step 3: Improve the network structure. First, three Coordinate Attention (CA) modules were added to the YOLOv7 network. Specifically, they were located in the YOLOv7 network head, after the ELAN feature extraction networks at layers 75, 88, and 101. This increased the network's sensitivity to target location information and improved its ability to detect multi-scale targets. Second, the feature fusion module in the original YOLOv7 network was replaced with the Spatial Pyramid Pooling–Fast (SPPF) module, reducing the number of parameters and computational complexity and accelerating particle morphology detection.

[0027] Step 4: Network training to obtain the optimal weights. The dataset annotated in step 2 and the improved network in step 3 are used for training. The hyperparameters are set as follows: the initial learning rate is set to 0.009, the learning rate momentum is set to 0.935, the weight decay coefficient is set to 0.0004, the cosine annealing parameter is set to 0.1, the bounding box loss weight coefficient is set to 0.05, the anchor scaling threshold is set to 4.0, and the IoU training threshold is set to 0.2. The training set divided in step 2 is input into the network training for 200 epochs to obtain the optimal weights. The morphological recognition results are shown in Figure 2. Figure 3 As shown in the figure, the morphologies of all particles can be accurately identified; moreover, the average accuracy of identification is as high as 86.4%, and the average detection delay is 15.7ms, achieving efficient detection of particle morphologies in mixed particle fields.

Claims

1. A particle morphology recognition method based on YOLOv7 and interferometric imaging technology, characterized in that: The method comprises the following steps: Step 1: Build an interferometric particle imaging system to obtain mixed particle interference defocus images. The interferometric particle imaging system includes: a laser, a spatial filter consisting of an objective lens and a pinhole, a collimating lens, a lens group consisting of a plano-convex cylindrical lens and a plano-concave cylindrical lens, a glass slide, an imaging lens, and a CCD camera. A 532 nm laser, a spatial filter consisting of an objective lens and a pinhole, a collimating lens, a lens group consisting of a plano-convex cylindrical lens and a plano-concave cylindrical lens, and a glass slide are arranged horizontally from left to right. An imaging lens and a CCD camera are arranged below the glass slide in sequence. Use this system to obtain a mixed particle field interference defocus image dataset: Sand and diluted polystyrene spherical particles are placed on a slide as the particles to be measured. After being irradiated by a sheet laser beam, a mixed particle field interference defocus image dataset is obtained at the CCD end. Step 2: Label and divide the dataset. Use LabelImg software to label the mixed particle field interference defocused image dataset obtained in step 1. The annotation content includes particle type and particle position information. After labeling, the mixed particle field interference defocused image dataset is divided into training set, validation set, and test set in a ratio of 7:1:

2. Step 3: Improve the network structure. First, add three collaborative attention modules (CA) to the YOLOv7 network. Specifically, they are located in the head of the YOLOv7 network, after the ELAN feature extraction network at layers 75, 88, and 101. Second, replace the feature fusion module in the original YOLOv7 network with a fast spatial pyramid pooling (SPPF) module, which reduces the number of parameters and computations and speeds up particle morphology detection. Step 4: Network training to obtain the optimal weight; use the dataset labeled in step 2 and the improved network in step 3 for training. The specific training steps are as follows: first set the hyperparameters. The hyperparameter settings are as follows: the initial learning rate is set to 0.009, the learning rate momentum is set to 0.935, the weight decay coefficient is set to 0.0004, the cosine annealing parameter is set to 0.1, the bounding box loss weight coefficient is set to 0.05, the anchor scaling threshold is set to 4.0, and the IoU training threshold is set to 0.

2. Secondly, the training set divided in step 2 is input into the network for 200 rounds of training to obtain the optimal weight; use the optimal weight to perform morphological recognition on the test set divided in step 2, and the average recognition accuracy reaches 86.4%, and the average detection delay is 15.7ms, achieving efficient detection of particle morphology in mixed particle fields.

2. The particle morphology recognition method based on YOLOv7 and interference imaging technology according to claim 1, characterized in that: 45 μm polystyrene spherical particles and sand particles are the particles to be tested and are placed on a glass slide.

3. The particle morphology recognition method based on YOLOv7 and interference imaging technology according to claim 1, characterized in that: A CCD camera with a resolution of 2448×2048 and a pixel size of 3.45 μm is used as the receiving device.

Images