Polyethylene inferior particle identification method and system, electronic device, and storage medium

Abstract

This invention discloses a method for identifying substandard polyethylene particles, used for identifying minute defects such as discolored spots on the surface and / or inside opaque (semi-transparent) particles. The method includes at least the following steps: A) Preprocessing and data augmenting images of defective polyethylene particles to obtain a polyethylene particle image set; B) Extracting multi-scale features from the defective polyethylene particle images and fusing these features to obtain a feature dataset; C) Constructing an initial identification model for minute defects in polyethylene particles and training the initial model based on the polyethylene particle image set to obtain a final identification model; D) Obtaining image features of the polyethylene particles to be detected under a single-layer rolling state on an inclined colored acrylic slide, and performing feature recognition of minute defects in the image features based on the final identification model. During feature recognition, a detection head based on discrete waveform transformation is used to enhance the pixel features of minute defects. This invention can significantly improve the false negative rate of polyethylene particles with minute defects both internally and externally.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method, system, electronic device, and storage medium for identifying inferior polyethylene particles. Background Technology

[0002] Polyethylene granules (also known as plastic granules) are an important raw material for injection molding and are widely used in various industries such as construction, automotive, communications, medical, and entertainment. Different products have different quality requirements for polyethylene granules. Defects such as black spots and discolored spots on the inside / surface of polyethylene granules are one of the important criteria for judging the quality of granule products, directly affecting the overall quality of the product and the economic benefits of polyethylene granule manufacturers.

[0003] Most polyethylene pellet manufacturers use visual sampling and inspection according to industry technical standards, manually identifying substandard pellets for quality control. Manual inspection suffers from numerous drawbacks, including high throughput, high false positive (or false negative) rates, difficulty in size calculations, and low efficiency. Furthermore, sampling inspection cannot comprehensively reflect the overall quality of a batch of products, failing to meet the urgent need for companies to improve product quality. Currently, technologies based on machine vision and intelligent algorithms for identifying and detecting substandard pellets have attracted attention. However, this technology is difficult to apply to particulate matter in different product sectors, lacks sufficient detection accuracy, and has not yet met the current practical requirements for intelligent quality inspection.

[0004] Chinese patent application CN112837311A discloses a deep learning-based polyethylene particle defect detection and identification system and method. The system includes a control unit, a feeding unit, an identification unit, a screening unit, and a collection component. The feeding unit causes the polyethylene particles to fall in a waterfall-like manner. The identification unit includes several multimodal cameras and an AI processor connected to each camera. Each multimodal camera is located on one side of the polyethylene particle's falling path and captures images of each particle, sending them to the AI ​​processor. The AI ​​processor is connected to the control unit and identifies defective polyethylene particles, sending these images to the control unit. While this deep learning-based polyethylene particle defect detection and identification system can achieve high throughput detection, the system and method perform image acquisition and analysis at high polyethylene particle speeds. It cannot effectively identify extremely small black spots on the particle surface or black or discolored spots inside the particles, resulting in a high overall false negative rate, which fails to meet actual production needs.

[0005] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a method for identifying inferior polyethylene particles, which is particularly suitable for identifying extremely small discolored defects on the surface of opaque or semi-transparent particles as well as discolored defects inside the particles. This method can greatly improve the false negative rate of polyethylene particles with extremely small defects on both the inside and outside.

[0007] To achieve the above objectives, according to a first aspect of the present invention, the present invention provides a method for identifying inferior polyethylene particles, used for identifying minute defects such as discolored dots on the surface and / or inside of opaque (semi-transparent) particles, comprising at least the following steps: A. Preprocessing and data augmenting images of defective polyethylene particles to obtain a polyethylene particle image set; B. Extracting multi-scale features from the defective polyethylene particle images and fusing the multi-scale features to obtain a feature dataset; C. Constructing an initial identification model for minute defects in polyethylene particles and training the initial identification model based on the polyethylene particle image set to obtain a final identification model; D. Obtaining image features of the polyethylene particles to be detected in a single-layer rolling state on an inclined colored acrylic slide, and performing feature recognition of minute defects in the image features based on the final identification model; During the feature recognition process, a detection head based on discrete waveform transformation is used to enhance the pixel features of minute defects.

[0008] Furthermore, in the above technical solution, discolored spots may include external black spots, internal black spots, and / or color grains. Minor defects refer to discolored spot defects of 5–20 pixels.

[0009] Furthermore, in the above technical solution, the colored acrylic slide can be made of transparent material and the surface of the slide is preferably frosted; the image of the polyethylene particles to be detected can be acquired by an industrial area scan camera under LED light board illumination.

[0010] Furthermore, in the above technical solution, step D, which uses a detection head based on discrete waveform transformation to enhance the pixel features of minute defects, can specifically include: D1, a convolutional detection head based on discrete waveform transformation, constructing three prediction layers for each detection head to predict target boxes at different scales; each prediction layer consists of a discrete waveform transformation module, a convolutional layer, and a fully connected layer; D2, position prediction using anchor boxes; in each prediction layer, a regression operation is used to match the anchor boxes to obtain the precise position of the target box; D3, in each prediction layer, the target category is predicted using a fully connected layer.

[0011] Furthermore, in the above technical solution, the discrete waveform transformation module can reduce the required computation by reducing the spatial dimension of the feature map; the convolutional layer can be used to extract features; and the fully connected layer can be used to generate the coordinates and class probabilities of the target box.

[0012] Furthermore, in the above technical solution, the identification method of the present invention may further include: E. further filtering and decoding the prediction results of the polyethylene particle image to be detected in step D, and generating identification detection results.

[0013] Furthermore, in the above technical solution, the preprocessing and data augmentation in step A may specifically include: A1. Obtaining a preprocessed image dataset through filtering, denoising, and sharpening enhancement; A2. Filtering multiple preprocessed images of particles with defects based on the defect categories of polyethylene particles; A3. Determining and labeling the typical geometric feature regions in each filtered image of particles with defects, and labeling each selected typical defect feature region with the defect type name according to the defect type to obtain an initial polyethylene particle image dataset; A4. Expanding the initial polyethylene particle image dataset using the Mosaic data augmentation algorithm to obtain a final polyethylene particle image dataset, which is used as both the training and testing datasets.

[0014] Furthermore, in the above technical solution, step B, extracting multi-scale features from the defective polyethylene particle image, can specifically include: B1, passing the polyethylene particle image to the backbone network as input and extracting image features through convolutional layers; B2, introducing a Residual Block structure into the backbone network to increase the network's depth and nonlinearity; B3, introducing CSP connections into the backbone network to improve feature representation and computational efficiency; B4, using a spatial pyramid pooling structure at the end of the backbone network to capture contextual information at different scales; B5, using a 1×1 convolutional layer to compress the number of channels in the feature map, reducing computation while retaining key feature information.

[0015] Furthermore, in the above technical solution, the feature fusion of multi-scale features in step B may specifically include: B6, generating feature maps of different scales through a feature pyramid network; B7, performing feature fusion using a path aggregation network.

[0016] Furthermore, in the above technical solution, the initial identification model for extremely small defects in polyethylene particles in step C can be constructed based on the backbone network optimized by Residual Block, the feature pyramid network, the path aggregation network, and the detection head network.

[0017] Furthermore, in the above technical solution, the filtering and decoding prediction results in step E can specifically include: E1, using the size and position information of the anchor boxes, decoding the target boxes output by the detection head network to obtain the true position of the target boxes; E2, using a non-maximum suppression algorithm to filter the decoded target boxes to reduce overlapping target boxes; E3, after filtering the target boxes, further filtering is performed based on the category probability of the target boxes; E4, the bounding boxes of the retained target boxes are adjusted to improve the accuracy of the detection results; E5, the filtered and adjusted target boxes are output as the final target detection results, and each target box contains at least the target's position, category, and confidence information.

[0018] According to a second aspect of the present invention, the present invention provides a polyethylene defect particle identification system for identifying minute defects of surface and / or internal discolored points on opaque (semi-transparent) particles, comprising at least: a particle image set acquisition module, which is used to preprocess and data augment images of polyethylene particles with defects to acquire a polyethylene particle image set; a feature dataset acquisition module, which is used to extract multi-scale features from the defective polyethylene particle images and perform feature fusion on the multi-scale features to acquire a feature dataset; an identification model acquisition module, which is used to construct an initial identification model for minute defects of polyethylene particles and train the initial identification model based on the polyethylene particle image set to acquire a final identification model; and a defect feature identification module, which is used to acquire image features of the polyethylene particles to be detected in a single-layer rolling state on an inclined colored acrylic slide and to perform feature identification of minute defects in the image features based on the final identification model; during the feature identification process, a detection head based on discrete waveform transformation is used to enhance the pixel features of minute defects.

[0019] Furthermore, in the above technical solution, the identification system of the present invention may further include: a detection result generation module, which is used to further filter and decode the prediction results of the image of the polyethylene particles to be detected, and generate identification detection results.

[0020] According to a third aspect of the present invention, the present invention provides an electronic device for identifying inferior polyethylene particles, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform a method for identifying inferior polyethylene particles as described in any of the above technical solutions.

[0021] According to a fourth aspect of the present invention, the present invention provides a non-transitory computer-readable storage medium storing computer-executable instructions for causing a computer to perform a method for identifying inferior polyethylene particles as described in any of the above technical solutions.

[0022] Compared with the prior art, the present invention has one or more of the following beneficial effects:

[0023] 1) By controlling the falling state of polyethylene particles, that is, by setting up colored acrylic slides and tilting the slides, the present invention creates a single-layer rolling state of the particles, which can effectively capture images of the particles to be detected from multiple angles and different particle surfaces. Compared with the "waterfall" falling of particles in the prior art, it can not only avoid missed detection to the greatest extent, but also make it easier to identify extremely small defects.

[0024] 2) In the defect feature recognition process, this invention uses a detection head based on discrete waveform transformation to enhance the pixel features of extremely small defects, which can effectively improve the resolution of extremely small defects. Therefore, the aforementioned control method of particle falling state and the method of enhancing the pixel features of extremely small defects have a synergistic effect. The combination of the two enables this invention to effectively identify 5-20 pixel color defects with high recognition efficiency.

[0025] 3) The present invention uses a transparent material for the colored acrylic slide and a frosted surface for the slide surface. This not only increases the friction of the rolling particles, thereby further enhancing the rolling effect of the particles, but the transparent material can also be combined with the LED light board to avoid the problem of image shadows when the industrial area scan camera takes pictures, which can further optimize the recognition effect of defective particles.

[0026] 4) This invention addresses the problem of identifying extremely small defects in polyethylene particles by creatively employing a convolutional detection head based on discrete waveform transform. For each detection head, three prediction layers are constructed to predict target boxes at different scales. Each prediction layer consists of a discrete waveform transform module (DWT), a convolutional layer, and a fully connected layer. DWT reduces the required computation by decreasing the spatial dimension of the feature map. Compared to convolution or pooling, DWT is lossless because it can reduce spatial resolution by expanding the number of channels by the same factor. Furthermore, DWT possesses additional compression properties that random filters or Fourier transforms lack.

[0027] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it according to the contents of the specification, and to make the above and other objects, technical features and advantages of the present invention easier to understand, one or more preferred embodiments are listed below and described in detail with reference to the accompanying drawings. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating the method for identifying inferior polyethylene particles according to the present invention.

[0029] Figure 2 This is a schematic diagram of the structure of the polyethylene inferior particle identification system of the present invention.

[0030] Figure 3 This is a schematic diagram of the hardware structure of the electronic device for identifying inferior polyethylene particles according to the present invention. Detailed Implementation

[0031] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings, but it should be understood that the scope of protection of the present invention is not limited to the specific embodiments.

[0032] Unless otherwise expressly stated, throughout the specification and claims, the term "comprising" or its variations such as "including" or "comprises" shall be understood to include the stated elements or components without excluding other elements or other components.

[0033] In this document, for ease of description, spatial relative terms such as “below,” “under,” “down,” “above,” “above,” “up,” etc., are used to describe the relationship of one element or feature to another element or feature in the accompanying drawings. It should be understood that spatial relative terms are intended to encompass different orientations of an object in use or operation, in addition to those depicted in the figures. For example, if an object in the figure is flipped, an element described as “below” or “under” another element or feature would be oriented “above” that element or feature. Thus, the exemplary term “below” can encompass both the downward and upward orientations. An object may also have other orientations (rotated 90 degrees or other orientations), and the spatial relative terms used herein should be interpreted accordingly.

[0034] In this document, the terms "first," "second," etc., are used to distinguish two different elements or parts, and are not used to define specific positions or relative relationships. In other words, in some embodiments, the terms "first," "second," etc., can also be used interchangeably.

[0035] The methods, systems, electronic devices, and storage media of the present invention are described in more detail below by way of specific embodiments. It should be understood that the embodiments are merely exemplary and the present invention is not limited thereto.

[0036] Example 1

[0037] like Figure 1 As shown, this embodiment provides a method for identifying inferior polyethylene particles, which can be used to identify extremely small defects such as discolored spots on the surface and / or inside opaque or semi-transparent particles. These extremely small defects refer to discolored spots of 5–20 pixels, which may include external black spots, internal black spots, and / or colored particles (dots). The identification method includes at least the following steps:

[0038] Step S100: Preprocess and data augment the images of polyethylene particles with defects to obtain a set of polyethylene particle images.

[0039] Step S200: Extract multi-scale features from the defective polyethylene particle image and perform feature fusion on the multi-scale features to obtain a feature dataset;

[0040] Step S300: Construct an initial identification model for extremely small defects in polyethylene particles, and train the initial identification model based on a set of polyethylene particle images to obtain the final identification model.

[0041] Step S400: Obtain the image features of the polyethylene particles to be detected in a single-layer rolling state on a tilted colored acrylic slide, and perform feature recognition on the minute defects in the image features based on the final recognition model; during the feature recognition process, a detection head based on discrete waveform transformation is used to enhance the pixel features of the minute defects. This step is a core step. Preferably, but not limitingly, the colored acrylic slide can be made of transparent material and the surface of the slide (i.e., the particle rolling surface) can be made of frosted material; the image of the polyethylene particles to be detected can be acquired by an industrial area scan camera under LED light panel illumination.

[0042] Step S500 is an optional step, which involves further filtering and decoding the prediction results of the polyethylene particle image to be detected in step S400, and generating the recognition and detection results.

[0043] The method described in this embodiment controls the falling state of polyethylene particles—specifically, by setting up a colored acrylic slide (preferably made of transparent acrylic, and the color can be dark blue) and tilting the slide—creating a single-layer rolling state of the particles. This allows for the effective capture of images of the particles to be detected from multiple angles and different particle surfaces. Compared to the "waterfall-like" falling of particles in existing technologies, this method not only minimizes missed detections but also makes it easier to identify extremely small defects. In the defect feature recognition process, this embodiment uses a detection head based on discrete waveform transformation to enhance the pixel features of extremely small defects, effectively improving the resolution of these defects. Therefore, the particle falling state control method and the method of enhancing the pixel features of extremely small defects in this embodiment have a synergistic effect. The combination of these two methods enables the present invention to effectively identify 5-20 pixel discoloration defects with high recognition efficiency.

[0044] In the aforementioned preferred solution, by using a transparent material for the colored acrylic slide and designing the slide surface to have a frosted surface, not only is the friction of the particles rolling off increased, thereby further enhancing the particle rolling effect, but the transparent material can also be combined with the supplementary lighting of the LED light panel to avoid the problem of image shadows when the industrial area scan camera takes pictures, which can further optimize the recognition effect of defective particles.

[0045] This embodiment preferably uses a convolutional neural network to construct the defect particle recognition model. Specifically, step S100 may include:

[0046] Step S101: Perform filtering, denoising, and sharpening enhancement processing on the image of imperfect (semi-)transparent polyethylene particles to obtain a preprocessed image dataset.

[0047] Step S102: Filter multiple preprocessed images containing defective particles based on the defect category of polyethylene particles.

[0048] Step S103: Identify the typical geometric feature regions in each filtered image of particles containing defects, and label the typical defect feature regions in each filtered particle image to obtain the initial particle image dataset. In one example, a selection method using a labeling tool can be used; in other examples, other methods that can identify typical geometric feature regions can also be used.

[0049] Step S104: Label each selected typical defect feature area with the defect type name according to the defect type to obtain the initial particle image dataset; the defect type may include discolored dots (i.e., external black dots, internal black dots and / or colored particles (dots), etc.), and the method of the present invention can also identify more easily identifiable defect types such as black particles, trailing particles, and snake-skin particles.

[0050] Step S105: Scale the images in the initial particle image dataset to a fixed size required by the model, preferably scaling the images to 256×256 pixels.

[0051] Step 106: Expand the initial particle image dataset. Specifically, the Mosaic data augmentation algorithm can be used to randomly stitch together four different images from the dataset to form a new sample that is then incorporated into the dataset, resulting in a final polyethylene particle image set that is augmented and conforms to the model input size. Specifically, three images can be randomly selected from the dataset and scaled, rotated, and flipped to increase sample diversity. Then, these three processed images are stitched together with the original image to form a new stitched image. On the stitched image, a region is randomly selected, and the pixel values ​​within this region are replaced with the pixel values ​​of the corresponding region in the fourth image, resulting in a new Mosaic data-augmented sample.

[0052] Step 107: Divide the final polyethylene particle image set into a training dataset and a test dataset. Specifically, in this embodiment, 70% of the data in the particle classification dataset can be used as the training set, and the remaining 30% as the test set.

[0053] Furthermore, step S200 may include:

[0054] Step S201: Pass the polyethylene particle image to the backbone network as input.

[0055] Step S202: Image Feature Extraction: The backbone network extracts image features through a series of convolutional layers. These convolutional layers can use 3x3 kernels and a stride of 2 to reduce the size of the feature maps.

[0056] Step S203: Introduce a Residual Block structure into the backbone network to increase the network's depth and non-linearity. The Residual Block consists of multiple convolutional layers and residual connections, where the residual connections directly add the input feature map to the output feature map, which can better propagate gradients and learn more complex features.

[0057] Step S204: Introduce CSP connections (Communication Sequential Processes) into the backbone network to improve feature representation and computational efficiency. CSP connections divide the feature map into two parts, one part is processed through a series of convolutional layers, and the other part is directly connected to the output to reduce computation and better propagate gradients and learn rich features.

[0058] Step S205: At the end of the backbone network, a Spatial Pyramid Pooling (SPP) structure is used to capture contextual information at different scales. The SPP module extracts features on pooling layers of different sizes and then concatenates these features to obtain a more global and multi-scale feature representation, which helps to improve the discriminative power of features and the recognition performance of images.

[0059] Step S206: Compress the number of channels in the feature map using a 1×1 convolutional layer to further reduce computation. This convolutional operation reduces computational complexity by lowering the dimensionality of the features while ensuring that key feature information is preserved.

[0060] Step S207: Generate feature maps at different scales using a Feature Pyramid Network (FPN). The FPN consists of a series of convolutional and upsampling layers, which extract feature pyramids of different scales from different layers of the backbone network. This pyramid structure ensures that higher-level semantic information is preserved at lower scales, while more detailed, fine-grained information is retained at higher scales.

[0061] Step S208: Feature fusion is performed using the Path Aggregation Network (PANet). PANet utilizes multi-scale information from the feature pyramid for feature fusion and contextual information transfer, and may include a feature upsampling module and a feature fusion module. The feature upsampling module upsamples low-resolution feature maps to the same size as high-resolution feature maps to maintain spatial consistency; methods such as nearest neighbor interpolation or bilinear interpolation can be used for upsampling. The feature fusion module achieves feature fusion by element-wise adding the upsampled low-resolution feature map to the corresponding high-resolution feature map. This fusion method preserves detailed information from different scales and improves the accuracy of object detection. Through residual connections, information from the low-resolution feature map can be added to the high-resolution feature map, making the fused features richer.

[0062] Furthermore, step S300 may specifically include:

[0063] Step S301: Construct an initial identification model for extremely small defects in opaque (semi-)transparent polyethylene particles based on the backbone network, feature pyramid network, and path aggregation network optimized by Residual Block, combined with the detection head network;

[0064] Step S302: Based on the training dataset, train the initial recognition model for extremely small defects in opaque (semi-)transparent polyethylene particles; use the particle images in the training dataset one by one in the initial order as input data for the initial model, calculate the initial model to obtain the input data, compare the output data with the particle type recognition results of the particle images, and correct the parameters of the initial model according to the comparison results; change the order of the particle images in the training dataset, and use the particle images in the training dataset with the changed order to test the corrected model again until the number of training times equals the preset number of times, and use the initial model with the last corrected weight parameters as the trained recognition model.

[0065] Step S303: Based on the test dataset, test the trained recognition model and output the final recognition model that meets the preset standards. Specifically, the trained recognition model can be verified using test data in the test dataset. The final recognition model is determined based on the accuracy of the model on the test dataset and the highest accuracy of the trained recognition model on the training dataset.

[0066] Furthermore, step S400 may specifically include:

[0067] Step S401: A convolutional detection head based on Discrete Waveform Transform (DWT) is constructed, with three prediction layers for each head to predict bounding boxes at different scales. Each prediction layer consists of a Discrete Waveform Transform (DWT) module, a convolutional layer, and a fully connected layer. DWT reduces the required computation by reducing the spatial dimension of the feature map. Unlike convolution or pooling, DWT is lossless because it can reduce spatial resolution by expanding the number of channels by the same factor. Furthermore, DWT has additional compression properties that random filters or Fourier transforms do not possess. Specifically, DWT first extracts learnable, spatially invariant features through convolutional layers, then performs spatial token mixing and downsampling to extract scale-invariant features using Discrete Two-Dimensional Wavelet Transform (2D-DWT). Since 2D-DWT is a lossless transform, it expands the number of channels by computing an approximate subband (low-resolution approximation) and three detail subbands (spatial derivative) for each input channel, thereby reducing spatial resolution. Next, a learnable MLP is used for channel mixing, consisting of two 1x1 convolutional layers separated by a GELU nonlinear function. Finally, a transposed convolutional layer restores the spatial resolution of the feature map. Convolutional layers are used for feature extraction, while fully connected layers generate the coordinates and class probabilities of the bounding boxes. Through layer-by-layer convolutional operations, the dimensionality of the features can be gradually reduced, and higher-level semantic information can be extracted.

[0068] Step S402: Predict location using anchor boxes. Define a set of predefined anchor boxes with different sizes and aspect ratios (D). In each prediction layer, use a regression operation to match the anchor boxes to obtain the precise location of the target box. The regression operation can predict the center coordinates, width, height, and other information of the target box, and is trained and optimized using an appropriate loss function.

[0069] Step S403: Predict the target class. Each prediction layer contains a fully connected layer for predicting the target class. The fully connected layer maps the feature map to the dimension of the number of classes and uses a softmax function to convert the output into class probabilities. A classification loss function is used to measure the difference between the predicted class and the true class, and to optimize the model's classification ability.

[0070] Furthermore, step S500 may specifically include:

[0071] Step S501: Using the size and position information of Anchor Boxes, the target boxes output by the detection head network can be decoded to obtain the true position of the target boxes. Based on the coordinate offset and scale transformation information in the prediction results output by the detection head network, the true position of the target boxes is calculated by combining them with the Anchor Boxes.

[0072] Step S502: Use Non-Maximum Suppression (NMS) to filter the decoded bounding boxes to reduce overlapping boxes. NMS filters boxes based on their confidence scores and overlap, retaining the boxes with the highest confidence scores and removing other boxes with high overlap. By adjusting the overlap threshold, the number of bounding boxes and the degree of overlap can be controlled.

[0073] Step S503: After filtering the target boxes using NMS, further filtering is performed based on the category probability of the target boxes. A category probability threshold is set, and only target boxes with a category probability higher than the threshold are retained. This further ensures that the target boxes belong to meaningful target categories and filters out target boxes with low confidence.

[0074] Step S504: For the retained target boxes, perform fine-tuning and adjustment operations on the bounding boxes to improve the accuracy of the detection results.

[0075] Step S505: Output the filtered and adjusted target bounding boxes as the final target detection results. Each target bounding box contains information such as the target's location, category, and confidence level.

[0076] Example 2

[0077] Combination Figure 2 As shown, this embodiment provides a polyethylene defect particle identification system for identifying minute defects such as discolored spots on the surface and / or inside opaque and semi-transparent particles. It includes at least a particle image set acquisition module 1, a feature dataset acquisition module 2, a recognition model acquisition module 3, and a defect feature recognition module 4. Specifically, the particle image set acquisition module 1 preprocesses and augments images of defective polyethylene particles to obtain a polyethylene particle image set; the feature dataset acquisition module 2 extracts multi-scale features from the defective polyethylene particle images and fuses these features to obtain a feature dataset; the recognition model acquisition module 3 constructs an initial recognition model for minute defects in polyethylene particles and trains the initial model based on the polyethylene particle image set to obtain a final recognition model; the defect feature recognition module 4 acquires image features of the polyethylene particles to be detected in a single-layer rolling state on an inclined colored acrylic slide and performs feature recognition of minute defects in the image features based on the final recognition model; during feature recognition, a detection head based on discrete waveform transformation is used to enhance the pixel features of the minute defects.

[0078] The identification system in this embodiment may further include a detection result generation module 5, which can be used to further filter and decode the image of the polyethylene particles to be detected and generate identification detection results.

[0079] The identification system in this embodiment is a virtual system corresponding to method embodiment 1, and can achieve the same technical effect as method embodiment 1.

[0080] Example 3

[0081] This embodiment provides a non-transient (non-volatile) computer storage medium storing computer-executable instructions. These instructions can execute the polyethylene defective particle identification method in any of the above-described method embodiments 1 and achieve the same technical effect. The identification method includes at least the following steps: A. Preprocessing and data augmenting images of defective polyethylene particles to obtain a polyethylene particle image set; B. Extracting multi-scale features from the defective polyethylene particle images and fusing these features to obtain a feature dataset; C. Constructing an initial identification model for extremely small defects in polyethylene particles and training the initial model based on the polyethylene particle image set to obtain a final identification model; D. Obtaining image features of the polyethylene particles to be detected in a single-layer rolling state on an inclined colored acrylic slide, and performing feature identification of extremely small defects in the image features based on the final identification model; During feature identification, a detection head based on discrete waveform transformation is used to enhance the pixel features of extremely small defects.

[0082] Example 4

[0083] This embodiment provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, which, when executed by a computer, cause the computer to perform the polyethylene inferior particle identification method described in the above aspects and achieve the same technical effect.

[0084] Example 5

[0085] Figure 3 This is a schematic diagram of the hardware structure of the polyethylene inferior particle identification electronic device according to this embodiment. The device includes one or more processors 610 and a memory 620. Taking one processor 610 as an example, the device may also include an input device 630 and an output device 640.

[0086] The processor 610, memory 620, input device 630 and output device 640 can be connected by a bus or other means.

[0087] The memory 620, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules. The processor 610 executes various functional applications and data processing of the electronic device by running the non-transitory software programs, instructions, and modules stored in the memory 620, thereby implementing the processing method of the above-described method embodiments.

[0088] The memory 620 may include a program storage area and a data storage area, wherein the program storage area may store the operating system and applications required for at least one function; the data storage area may store data, etc. Furthermore, the memory 620 may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory 620 may optionally include memory remotely located relative to the processor 610, and these remote memories may be connected to the processing device via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0089] Input device 630 can receive input digital or character information and generate signal input. Output device 640 may include display devices such as a display screen.

[0090] The one or more modules are stored in the memory 620. When executed by the one or more processors 610, they perform the following: A. Preprocessing and data augmentation of the polyethylene particle images with defects to obtain a polyethylene particle image set; B. Extracting multi-scale features from the defective polyethylene particle images and fusing the multi-scale features to obtain a feature dataset; C. Constructing an initial recognition model for extremely small defects in polyethylene particles and training the initial recognition model based on the polyethylene particle image set to obtain a final recognition model; D. Obtaining the image features of the polyethylene particles to be detected in a single-layer rolling state on an inclined colored acrylic slide, and performing feature recognition of extremely small defects in the image features based on the final recognition model; During the feature recognition process, a detection head based on discrete waveform transformation is used to enhance the pixel features of extremely small defects.

[0091] The above-described product can execute the methods provided in the embodiments of the present invention, and has the corresponding functional modules and beneficial effects for executing the methods. Technical details not described in detail in this embodiment can be found in the methods provided in other embodiments of the present invention.

[0092] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0093] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0094] The foregoing description of specific exemplary embodiments of the present invention is for illustrative and explanatory purposes. These descriptions are not intended to limit the invention to the precise forms disclosed, and it will be apparent that many changes and variations can be made in accordance with the foregoing teachings. The exemplary embodiments were chosen and described in order to explain the specific principles of the invention and its practical application, thereby enabling those skilled in the art to implement and utilize various different exemplary embodiments of the invention, as well as various different choices and variations. Any simple modifications, equivalent changes, and alterations made to the foregoing exemplary embodiments should fall within the scope of protection of the present invention.

Claims

1. A method for identifying inferior polyethylene particles, characterized in that, For the identification of minute defects such as discolored dots on the surface and / or inside of opaque or semi-transparent particles, the following steps are included: A. Preprocess and data augment the images of polyethylene particles with defects to obtain a set of polyethylene particle images; B. Extract multi-scale features from images of defective polyethylene particles and perform feature fusion on the multi-scale features to obtain a feature dataset; C. Construct an initial identification model for extremely small defects in polyethylene particles, and train the initial identification model based on the polyethylene particle image set to obtain the final identification model; D. Obtain the image features of the polyethylene particles to be detected in a single-layer rolling state on a tilted colored acrylic slide, and perform feature recognition on the minute defects in the image features based on the final recognition model. During the feature recognition process, a detection head based on discrete waveform transformation is used to enhance the pixel features of the extremely small defects.

2. The method for identifying inferior polyethylene particles according to claim 1, characterized in that, The discolored dots include external black dots, internal black dots, and / or color grain dots.

3. The method for identifying inferior polyethylene particles according to claim 1, characterized in that, The extremely minor defects are 5 to 20 pixels of off-color spots.

4. The method for identifying inferior polyethylene particles according to claim 1, characterized in that, The colored acrylic slide is made of transparent material and the surface of the slide is frosted; the image of the polyethylene particles to be detected is acquired by an industrial area scan camera under LED light board illumination.

5. The method for identifying inferior polyethylene particles according to claim 1, characterized in that, The step D, which uses a detection head based on discrete waveform transformation to enhance the pixel features of the minute defects, specifically includes: D1. A convolutional detection head based on discrete waveform transformation is constructed for each detection head, which is used to predict target boxes at different scales. Each prediction layer consists of a discrete waveform transformation module, a convolutional layer and a fully connected layer. D2. Predict the location using anchor boxes; in each prediction layer, use a regression operation to match the anchor boxes to obtain the precise location of the target box; D3. In each prediction layer, the target category is predicted through the fully connected layer.

6. The method for identifying inferior polyethylene particles according to claim 5, characterized in that, The discrete waveform transformation module reduces the required computation by reducing the spatial dimension of the feature map; the convolutional layer is used to extract features; and the fully connected layer is used to generate the coordinates and class probabilities of the target box.

7. The method for identifying inferior polyethylene particles according to claim 1, characterized in that, The identification method further includes: E. Further filter and decode the image of the polyethylene particles to be detected in step D, and generate the identification and detection results.

8. The method for identifying inferior polyethylene particles according to claim 1, characterized in that, The preprocessing and data augmentation in step A specifically include: A1. Obtain the preprocessed image dataset through filtering, denoising, and sharpening enhancement. A2. Filter multiple preprocessed images containing defective particles based on the defect category of polyethylene particles; A3. Identify and label the typical geometric feature regions in each filtered image of defective particles. Label each selected typical defect feature region with the defect type name according to the defect type to obtain the initial polyethylene particle image dataset. A4. The initial polyethylene particle image dataset is augmented using the Mosaic data augmentation algorithm to obtain the final polyethylene particle image dataset, which is then used as the training and testing datasets.

9. The method for identifying inferior polyethylene particles according to claim 1, characterized in that, The extraction of multi-scale features from the defective polyethylene particle image in step B specifically includes: B1. The polyethylene particle image is passed to the backbone network as input, and image features are extracted through convolutional layers. B2. Introduce a ResidualBlock structure into the backbone network to increase the network's depth and nonlinearity. B3. Introduce CSP connections into the backbone network to improve the representation capability and computational efficiency of features; B4. A spatial pyramid pooling structure is used at the end of the backbone network to capture contextual information at different scales. B5. Use a 1×1 convolutional layer to compress the number of channels in the feature map, reducing computation while preserving key feature information.

10. The method for identifying inferior polyethylene particles according to claim 9, characterized in that, Step B, which involves feature fusion of multi-scale features, specifically includes: B6. Generate feature maps of different scales using a feature pyramid network; B7. Use path aggregation networks for feature fusion.

11. The method for identifying inferior polyethylene particles according to claim 10, characterized in that, The initial identification model for extremely small defects in polyethylene particles in step C is constructed based on a backbone network optimized by ResidualBlock, a feature pyramid network, a path aggregation network, and a detection head network.

12. The method for identifying inferior polyethylene particles according to claim 7, characterized in that, The filtering and decoding prediction results in step E specifically include: E1. Using the size and position information of the anchor frame, the target box output by the detection head network is decoded to obtain the true position of the target box; E2. Use the non-maximum suppression algorithm to filter the decoded target boxes to reduce overlapping target boxes; E3. After filtering the target boxes, further filtering is performed based on the category probability of the target boxes; E4. Adjust the bounding box of the retained target box to improve the accuracy of the detection results; E5. Output the filtered and adjusted bounding boxes as the final target detection results. Each bounding box should contain at least the target's location, category, and confidence information.

13. A system for identifying inferior polyethylene particles, characterized in that, Used for identifying minute defects such as discolored dots on the surface and / or inside of opaque or translucent particles, including: The particle image set acquisition module is used to preprocess and data augment images of polyethylene particles with defects to acquire a polyethylene particle image set. The feature dataset acquisition module is used to extract multi-scale features from images of defective polyethylene particles and perform feature fusion on the multi-scale features to obtain the feature dataset. The identification model acquisition module is used to construct an initial identification model for extremely small defects in polyethylene particles, and to train the initial identification model based on the polyethylene particle image set to obtain the final identification model. The defect feature recognition module is used to acquire the image features of the polyethylene particles to be detected in a single-layer rolling state on a tilted colored acrylic slide, and to perform feature recognition of the extremely small defects in the image features based on the final recognition model; during the feature recognition process, a detection head based on discrete waveform transformation is used to enhance the pixel features of the extremely small defects.

14. The polyethylene inferior particle identification system according to claim 13, characterized in that, The identification system also includes: The detection result generation module is used to further filter and decode the image of the polyethylene particles to be detected, and generate the identification detection result.

15. An electronic device for identifying inferior polyethylene particles, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to cause the at least one processor to perform the polyethylene inferior particle identification method as described in any one of claims 1 to 12.

16. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer-executable instructions for causing the computer to perform the polyethylene inferior particle identification method as described in any one of claims 1 to 12.

Images