A PBAT degradation film material type identification method and system based on image processing technology

Abstract

The present application discloses a method and system for identifying the type of PBAT degradation film material based on image processing technology, wherein the method includes performing shadow compensation on the initial optical image of the target degradation film material taken by the optical imaging device to obtain the real optical image of the target degradation film material; performing block processing on the real optical image by an adaptive algorithm based on the image content, extracting each pixel in each image block as a graph node, and constructing graph structure data; performing feature encoding on the graph structure data based on the graph attention network to obtain the encoded feature vector, decoding the feature vector, and obtaining the decoded feature map; constructing a corresponding number of recognition models based on the type of degradation film material; inputting the feature map into each recognition model to obtain the matching degree between the target degradation film material and each degradation film material type, and identifying the type of the target degradation film material according to the matching degree. The method provided in the present application can automatically identify the type of PBAT degradation film material.

Description

Technical Field

[0001] The present application relates to the field of image recognition, and in particular to a method and system for identifying the type of PBAT degradation film material based on image processing technology. Background Art

[0002] PBAT is a biodegradable plastic with the advantages of being biodegradable, environmentally friendly, and recyclable. It is widely used in various fields, such as food packaging, agriculture, medical devices, etc.

[0003] In actual production, it is often necessary to classify different types of PBAT degradation film material products. In the past, manual classification had a large error rate, resulting in the final products being affected by substandard performance, easy breakage, poor degradation performance, etc. Summary of the invention

[0004] The present application aims to provide a method and system for identifying the type of PBAT degradation film materials based on image processing technology to solve the technical problem of inaccurate identification of the type of PBAT degradation film materials.

[0005] In order to solve the above technical problems, the embodiment of the present application provides a method for identifying the type of PBAT degradation film material based on image processing technology, including:

[0006] Performing shadow compensation on an initial optical image of a target degradation film material taken by an optical imaging device to obtain a true optical image of the target degradation film material, wherein the shadow compensation is designed to eliminate the influence of the missing depth value of a transparent object on the recognition accuracy of the image of the target degradation film material;

[0007] The real optical image is processed by a block-by-block process using an adaptive algorithm based on image content to obtain a plurality of image blocks;

[0008] Extracting each pixel point in each of the image blocks as a graph node, calculating the weight of an edge according to the adjacency relationship and feature similarity between any two of the pixel points, and constructing graph structure data according to the graph nodes and the weights of the edges;

[0009] Based on the graph attention network, feature encoding is performed on the graph structure data to obtain an encoded feature vector, and the feature vector is decoded to obtain a decoded feature graph;

[0010] Constructing a corresponding number of recognition models based on the type of degradation film material, wherein each recognition model corresponds to one type of degradation film material;

[0011] The characteristic graphs are input into the respective identification models to obtain the matching degree between the target degradation film material and the types of the degradation film materials, and the type of the target degradation film material is identified according to the matching degree.

[0012] As one preferred solution, performing shadow compensation on an initial optical image of a target degradable film material taken by an optical imaging device to obtain a real optical image of the target degradable film material comprises:

[0013] Inputting the initial optical image into a pre-trained shadow detection model to perform shadow detection, thereby obtaining a shadow region of the initial optical image, wherein the shadow detection model is designed to correct a prediction result of the shadow detection model through real-time shadow error detection;

[0014] Processing the initial optical image according to the shadow region and the shadow removal algorithm to obtain a shadow-free image;

[0015] Extracting the depth value of the initial optical image to obtain the depth information of the target degradation film material;

[0016] The shadow-free image and the depth information are processed based on a stereo matching algorithm to obtain a true optical image of the target degradation film material.

[0017] As one preferred solution, the processing of the initial optical image according to the shadow area and the shadow removal algorithm to obtain a shadow-free image includes:

[0018] Converting the initial optical image into the HSV color space to obtain an HSV image;

[0019] Distinguish the shadow area and the non-shadow area by the brightness value of the HSV image;

[0020] The brightness value of the shadow area is corrected based on the brightness value of the non-shadow area, which is expressed as:

[0021] ;

[0022] in, represents the corrected HSV image, represents the HSV image before correction, and is the adjustment factor, is the brightness threshold;

[0023] The corrected HSV image is converted into the RGB color space to obtain the shadow-free image.

[0024] As one preferred solution, the extracting the depth value of the initial optical image to obtain the depth information of the target degradation film material includes:

[0025] Dividing the space where the initial optical image is located into a voxel grid;

[0026] Simulating rays emitted by an imaging device based on a ray tracing algorithm, and calculating intersection points of each ray with the voxel grid according to an intersection detection algorithm;

[0027] The characteristic information of the intersection is obtained, and the characteristic information is input into a pre-trained depth prediction model to obtain the depth information of the target degradation film material.

[0028] As one preferred solution, before inputting the feature information into the pre-trained depth prediction model, the method further includes:

[0029] Collecting an RGB-D image data set of the degradation film material, wherein the RGB-D image data set includes depth information of the degradation film material;

[0030] Dividing the space where the RGB-D image data set is located into a voxel grid, extracting target features from the RGB-D image data set, mapping the target features to the voxel space, generating a voxel feature vector for each voxel, and obtaining a voxel feature vector set;

[0031] A basic depth prediction model based on a deep convolutional neural network is constructed, and the basic depth prediction model is trained based on the voxel feature vector set to obtain a depth prediction model.

[0032] As one of the preferred solutions, the real optical image is processed into blocks by an adaptive algorithm based on image content to obtain multiple image blocks, including:

[0033] Preprocessing the real optical image, wherein the preprocessing includes denoising and contrast enhancement;

[0034] Extracting feature information of the preprocessed real optical image based on image processing technology, wherein the feature information includes at least spatial features, texture features and color features;

[0035] Calculating the similarity between different regions of the real optical image according to the feature information;

[0036] Based on the calculation result of the similarity, the real optical image is divided into a plurality of image blocks.

[0037] As one of the preferred solutions, extracting each pixel point in each of the image blocks as a graph node, calculating the weight of the edge according to the adjacency relationship and feature similarity between any two of the pixel points, and constructing graph structure data according to the graph nodes and the weights of the edges, includes:

[0038] In each image block, each pixel point is characterized as a graph node, and the graph node at least includes the spatial feature, texture feature and color feature of the pixel point;

[0039] Determine the adjacency relationship between any two of the graph nodes based on the spatial features, calculate the feature similarity between any two of the pixel points based on the texture features and the color features, and calculate the edge weight corresponding to each of the graph nodes based on the adjacency relationship and the feature similarity;

[0040] Graph structure data is constructed according to the graph nodes and the edge weights.

[0041] As one of the preferred solutions, the feature encoding of the graph structure data based on the graph attention network to obtain the encoded feature vector, and decoding the feature vector to obtain the decoded feature map includes:

[0042] Inputting the constructed graph structure data into the graph attention network, iteratively updating the nodes in the graph structure data through the multi-layer attention mechanism of the graph attention network, and obtaining the feature vector representation of each node;

[0043] The feature vector of each node is converted by a decoder to obtain a decoded feature map.

[0044] As one preferred solution, the characteristic graph is input into each of the identification models to obtain the matching degree between the target degradation film material and each type of degradation film material, and the type of the target degradation film material is identified according to the matching degree, including:

[0045] Inputting the characteristic graph of the target degradation film material into each of the identification models respectively, and obtaining the matching degree between the target degradation film material and the corresponding degradation mold material type, wherein the matching degree reflects the similarity between the target degradation film material and the corresponding degradation mold material;

[0046] The matching degrees outputted by the various recognition models are compared, and the type of degradation film material corresponding to the model with the highest matching degree is selected as the type of the target degradation film material.

[0047] Another embodiment of the present application provides a PBAT degradation film material type identification system based on image processing technology, including:

[0048] A compensation module, used for performing shadow compensation on an initial optical image of a target degradation film material taken by an optical imaging device to obtain a real optical image of the target degradation film material, wherein the shadow compensation is designed to eliminate the influence of the missing depth value of a transparent object on the recognition accuracy of the image of the target degradation film material;

[0049] A partitioning module, configured to perform a partitioning process on the real optical image by using an adaptive algorithm based on image content to obtain a plurality of image blocks;

[0050] An extraction module, used to extract each pixel point in each of the image blocks as a graph node, calculate the weight of the edge according to the adjacency relationship and feature similarity between any two of the pixel points, and construct graph structure data according to the graph nodes and the weights of the edges;

[0051] An encoding module, used for performing feature encoding on the graph structure data based on a graph attention network to obtain an encoded feature vector, and decoding the feature vector to obtain a decoded feature graph;

[0052] A construction module, used to construct a corresponding number of recognition models based on the type of degradation film material, wherein each of the recognition models corresponds to one type of degradation film material;

[0053] The identification module is used to input the characteristic graph into each of the identification models respectively, obtain the matching degree between the target degradation film material and each type of degradation film material, and identify the type of the target degradation film material according to the matching degree.

[0054] Compared with the prior art, the beneficial effects of the embodiments of the present application are at least one of the following:

[0055] (1) This application eliminates the impact of the lack of depth value of transparent objects on image recognition accuracy through shadow compensation technology. As a transparent or translucent object, PBAT degradation film material is easily affected by lighting conditions during the imaging process, resulting in shadows or reflections in the image. Shadow compensation technology can effectively reduce these interference factors, making the image more realistically reflect the characteristics of the material, thereby improving the accuracy of recognition.

[0056] (2) This application uses an adaptive algorithm based on image content to process real optical images in blocks, which can be flexibly divided according to the feature differences of different regions in the image, so that each image block can better reflect the local characteristics of the material. This processing method helps to capture more detailed information and further improve the accuracy of recognition.

[0057] (3) This application uses a graph attention network to encode and decode features of graph structure data, which can automatically learn the complex relationships between pixels in the image and highlight key features. This method can more accurately capture the unique texture, color and other features of PBAT degradation film materials, thereby improving recognition accuracy.

[0058] (4) Compared with manual classification methods, this application can significantly reduce the classification error rate. This helps to reduce additional costs such as rework and scrapping of finished products caused by inaccurate classification. Accurate identification methods help ensure that the quality of PBAT degradation film materials meets standard requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a schematic diagram of a process of identifying the type of PBAT degradation film material based on image processing technology in one embodiment of the present application;

[0060] Figure 2 is a schematic diagram of a shadow detection model in one embodiment of the present application;

[0061] Figure 3 It is a structural schematic diagram of a PBAT degradation film material type identification system based on image processing technology in one embodiment of the present application. DETAILED DESCRIPTION

[0062] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. The purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0063] In the description of this application, the terms "first", "second", "third", etc. are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first", "second", "third", etc. may explicitly or implicitly include one or more of the feature. In the description of this application, unless otherwise specified, "plurality" means two or more.

[0064] In the description of the present application, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be a connection between the two elements. The terms "vertical", "horizontal", "left", "right", "upper", "lower" and similar expressions used in this article are only for illustrative purposes, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on this application. The term "and / or" used in this article includes any and all combinations of one or more related listed items. For those of ordinary skill in the art, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0065] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this application have the same meaning as those commonly understood by those skilled in the art. The terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application. For those of ordinary skill in the art, the specific meanings of the above terms in this application can be understood by specific circumstances.

[0066] An embodiment of the present application provides a method for identifying the type of PBAT degradation film material based on image processing technology. For details, see Figure 1 , Figure 1 The figure shows a flow chart of a method for identifying the type of PBAT degradation film material based on image processing technology provided in one embodiment of the present application. It includes steps S1 to S6:

[0067] S1, performing shadow compensation on an initial optical image of a target degradation film material taken by an optical imaging device to obtain a true optical image of the target degradation film material, wherein the shadow compensation is designed to eliminate the influence of the missing depth value of a transparent object on the recognition accuracy of the image of the target degradation film material;

[0068] S2, performing block processing on the real optical image by an adaptive algorithm based on image content to obtain a plurality of image blocks;

[0069] S3, extracting each pixel point in each of the image blocks as a graph node, and calculating the weight of the edge according to the adjacency relationship and feature similarity between any two of the pixel points, and constructing graph structure data according to the graph nodes and the weights of the edges;

[0070] S4, performing feature encoding on the graph structure data based on a graph attention network to obtain an encoded feature vector, and decoding the feature vector to obtain a decoded feature graph;

[0071] S5, constructing a corresponding number of recognition models based on the type of the degradation film material, wherein each of the recognition models corresponds to one type of degradation film material;

[0072] S6. Input the characteristic graph into each of the identification models respectively to obtain the matching degree between the target degradation film material and each type of degradation film material, and identify the type of the target degradation film material according to the matching degree.

[0073] It should be noted in advance that PBAT degradable film materials are widely used in food packaging, agriculture, medical devices and other fields due to their advantages such as biodegradability, environmental protection and recyclability. However, different application scenarios have different performance requirements for PBAT degradable film materials, such as thickness, strength, degradation rate, etc. Therefore, there are many types of PBAT degradable film materials on the market to meet the needs of different customers. In the actual production process, in order to ensure product quality and meet customer needs, different types of PBAT degradable film materials need to be accurately classified. Traditional manual classification methods are not only time-consuming and labor-intensive, but also prone to errors, resulting in inaccurate classification, affecting subsequent production and processing processes. Therefore, an efficient and accurate classification method is needed to replace manual classification.

[0074] As a non-contact automatic recognition method, image recognition technology has the advantages of high efficiency, accuracy, and high degree of automation. Through image processing technology, the images of PBAT degradation film materials can be preprocessed, feature extracted, classified and recognized, so as to achieve accurate classification of different types of PBAT degradation film materials. This method not only improves the accuracy and efficiency of classification, but also reduces labor and time costs. However, as a transparent or translucent object, PBAT degradation film materials are easily affected by lighting conditions during the imaging process, resulting in shadows or reflections in the image. These interference factors will affect the accuracy of image recognition, making it difficult for traditional image recognition methods to be directly applied to the classification and recognition of PBAT degradation film materials.

[0075] Preferably, in one embodiment of the present invention, performing shadow compensation on an initial optical image of the target degradation film material taken by an optical imaging device to obtain a real optical image of the target degradation film material comprises:

[0076] Inputting the initial optical image into a pre-trained shadow detection model to perform shadow detection, thereby obtaining a shadow region of the initial optical image, wherein the shadow detection model is designed to correct a prediction result of the shadow detection model through real-time shadow error detection;

[0077] Processing the initial optical image according to the shadow region and the shadow removal algorithm to obtain a shadow-free image;

[0078] Extracting the depth value of the initial optical image to obtain the depth information of the target degradation film material;

[0079] The shadow-free image and the depth information are processed based on a stereo matching algorithm to obtain a true optical image of the target degradation film material.

[0080] The initial optical image is input into the pre-trained shadow detection model. This model is built based on deep learning technology and can accurately identify shadow areas in the image. At the same time, the model also has a real-time shadow error detection function, which can correct the prediction results to improve the accuracy of shadow detection. Specifically, the initial optical image of the target degradation film material taken by the optical imaging device is used as input. The image may contain shadow areas, and the pre-trained shadow detection model is used to perform shadow detection on the input image. This model is built based on deep learning technology and can identify and locate shadow areas in images.

[0081] like Figure 2 As shown in the figure, the shadow detection model includes a shadow feature extraction network based on the attention mechanism and a self-correcting deep learning network. In computer vision tasks, shadow detection requires accurate extraction of shadow features in images. The attention mechanism extracts the most significant features of the target foreground by assigning weights to pixel-level semantics, guiding the algorithm to focus on processing these primary target features, reducing the correlation between the foreground and the background, and thus improving the iteration speed and accuracy of the model.

[0082] In the actual shadow detection network, the ability to extract shadow features is enhanced by adding attention modules. These attention modules can be embedded in different layers of CNN to capture shadow features of different scales and levels. In order to make full use of feature information at different levels, the network also adopts a feature fusion strategy. In the downsampling and upsampling feature fusion stage, a feature map fusion algorithm based on the attention mechanism is designed to strengthen the data flow of shallow and deep features, thereby improving the accuracy of shadow detection.

[0083] The Self-Calibrated Network (SCNet) introduces the concept of self-calibrated convolutions, which aims to expand the field of view of each convolution layer through an internal communication mechanism to enrich the output features. Traditional convolution operations may limit the model's ability to capture long-term spatial dependencies and cross-channel relationships, while self-calibrated convolutions can adaptively construct long-distance spatial dependencies and inter-channel dependencies around each spatial position, enhancing the model's ability to integrate rich information. The self-calibrated deep learning network integrates self-calibrated convolution layers on the basic CNN architecture. These self-calibrated convolution layers can adaptively adjust the weights of the convolution kernel to better capture the shadow features in the image. At the same time, the network also adopts a deep supervision strategy to ensure accurate shadow feature representation at different levels.

[0084] The shadow feature extraction network based on the attention mechanism and the self-correction deep learning network are integrated to form a complete shadow detection model. The model can accurately identify and locate the shadow area in the image, and provide rich shadow feature information. In practical applications, the model can be used in the shadow compensation task of the initial optical image taken by the optical imaging device. By eliminating or reducing the shadow interference, an optical image that more realistically reflects the characteristics of the target degradation film material can be obtained.

[0085] Preferably, in one embodiment of the present invention, the processing of the initial optical image according to the shadow area and the shadow removal algorithm to obtain a shadow-free image comprises:

[0086] Converting the initial optical image into the HSV color space to obtain an HSV image;

[0087] Distinguish the shadow area and the non-shadow area by the brightness value of the HSV image;

[0088] The brightness value of the shadow area is corrected based on the brightness value of the non-shadow area, which is expressed as:

[0089] ,

[0090] in, represents the corrected HSV image, represents the HSV image before correction, and is the adjustment factor, is the brightness threshold;

[0091] The corrected HSV image is converted into the RGB color space to obtain the shadow-free image.

[0092] Specifically, use image processing software or library (such as OpenCV) to read the initial optical image taken by the optical imaging device, and convert the read initial optical image from the RGB color space to the HSV color space. Extract the brightness component (Value) from the HSV image. The brightness component reflects the brightness of the image and is the key to distinguishing between shadow areas and non-shadow areas. Set a suitable brightness threshold according to the actual situation of the image and the need for shadow correction. The brightness threshold is used to distinguish between shadow areas (lower brightness) and non-shadow areas (higher brightness). Compare the brightness component with the brightness threshold, and the area below the threshold is determined as a shadow area, and the area above the threshold is determined as a non-shadow area.

[0093] The brightness values ​​of the non-shadow area are counted and their average brightness value is calculated. The average brightness value is used as a reference for correcting the brightness value of the shadow area. According to actual needs, two adjustment coefficients are set. The adjustment coefficients are used to control the degree of correction of the brightness value of the shadow area. Usually, these two coefficients can be determined through experiments. According to the correction formula, the brightness value of the shadow area is corrected. The corrected brightness value should be close to the average brightness value of the non-shadow area to eliminate the influence of the shadow on the image.

[0094] Preferably, in one embodiment of the present invention, extracting the depth value of the initial optical image to obtain the depth information of the target degradation film material includes:

[0095] Dividing the space where the initial optical image is located into a voxel grid;

[0096] Simulating rays emitted by an imaging device based on a ray tracing algorithm, and calculating intersection points of each ray with the voxel grid according to an intersection detection algorithm;

[0097] The characteristic information of the intersection is obtained, and the characteristic information is input into a pre-trained depth prediction model to obtain the depth information of the target degradation film material.

[0098] Specifically, we first need to clarify the three-dimensional space range represented by the initial optical image, which is usually determined based on the objects and scenes in the image. Within the spatial range, the space is divided into a number of voxels (i.e., three-dimensional pixels) according to a certain size and resolution. Each voxel represents a small three-dimensional space area and has unique coordinates. For each voxel, it can be assigned certain attribute values ​​(such as density, refractive index, etc.), which will be used for subsequent ray tracing and depth prediction.

[0099] Furthermore, the parameters of the imaging device are determined, including the position, field of view, focal length and other parameters of the imaging device, which will affect the emission direction and range of the rays. Starting from each pixel point of the imaging device, rays are emitted according to the determined parameters. These rays will pass through the three-dimensional space and intersect with the voxel grid. For each ray, its emission point, direction, length and other information need to be recorded for subsequent intersection detection and depth prediction.

[0100] Common intersection detection algorithms include ray-triangle intersection detection, ray-voxel intersection detection, etc. In this scheme, since the space has been divided into voxel grids, the ray-voxel intersection detection algorithm can be selected. For each ray, an intersection detection algorithm is used to calculate its intersection with the voxel grid. This usually involves solving the ray equation and the geometric representation of the voxel grid to find the coordinates of the intersection of the ray and the voxel. For each intersection point, its coordinates, the voxel to which it belongs, and the light information at the time of intersection (such as direction, intensity, etc.) need to be recorded.

[0101] For each intersection, the voxel attribute values ​​around it (such as density, refractive index, etc.) and the light information at the time of intersection (such as direction, intensity, color, etc.) can be extracted as feature information. The extracted feature information is preprocessed, such as normalization, dimensionality reduction, etc., to improve the accuracy and efficiency of the depth prediction model. The preprocessed feature information is input into the pre-trained depth prediction model. This model is usually a neural network that has been trained with a large amount of data and can predict the depth information of the target degradation film material based on the input feature information.

[0102] Preferably, in one embodiment of the present invention, before inputting the feature information into the pre-trained depth prediction model, it also includes:

[0103] Collecting an RGB-D image data set of the degradation film material, wherein the RGB-D image data set includes depth information of the degradation film material;

[0104] Dividing the space where the RGB-D image data set is located into a voxel grid, extracting target features from the RGB-D image data set, mapping the target features to the voxel space, generating a voxel feature vector for each voxel, and obtaining a voxel feature vector set;

[0105] A basic depth prediction model based on a deep convolutional neural network is constructed, and the basic depth prediction model is trained based on the voxel feature vector set to obtain a depth prediction model.

[0106] Specifically, information such as the type, shape, size, and shooting environment (such as lighting conditions, shooting angle, etc.) of the degradation film material is collected. Use a camera that supports depth information acquisition (such as an RGB-D camera) to shoot the degradation film material. RGB-D images contain both color information (RGB) and depth information (D), where the depth information represents the distance from each pixel in the image to the camera. The captured RGB-D images are preprocessed, including denoising, correction, cropping, etc., to improve data quality. The preprocessed RGB-D images are organized according to certain rules (such as shooting time, material type, etc.) to construct a complete RGB-D image dataset.

[0107] According to the resolution of the RGB-D image and the size of the degradation film material, the size and resolution of the voxel grid are determined. The size of the voxel grid should be able to accurately reflect the structural characteristics of the degradation film material. In the three-dimensional space where the RGB-D image dataset is located, it is divided according to the determined voxel grid size to obtain a number of voxels. Features related to the target degradation film material are extracted from the RGB-D image, such as color features (RGB values), texture features, shape features, etc. At the same time, the depth information is used to extract spatial position features and depth change features. The extracted target features are mapped to the voxel grid, and a voxel feature vector containing the target features is generated for each voxel. The voxel feature vector should be able to accurately reflect the material characteristics and spatial structure of the location of the voxel. The feature vectors of all voxels are combined together to construct a voxel feature vector set. This set will serve as the input data for the subsequent depth prediction model.

[0108] According to the task requirements and data characteristics, a suitable deep convolutional neural network is selected as the basic model. The deep convolutional neural network has powerful feature extraction and learning capabilities and is suitable for processing complex image data. Based on the selected deep convolutional neural network, a basic depth prediction model is constructed. The model should include input layer, feature extraction layer, fully connected layer and output layer. Among them, the input layer is used to receive the voxel feature vector set; the feature extraction layer is used to extract the key information in the voxel features; the fully connected layer is used to integrate the feature information into the depth prediction result; and the output layer outputs the predicted depth information. The basic depth prediction model is trained using the voxel feature vector set and the corresponding depth information. During the training process, the parameters and structure of the model are adjusted so that the model can accurately predict the depth information of the degradation film material. After the training is completed, a trained depth prediction model is obtained.

[0109] Preferably, in an embodiment of the present invention, the real optical image is processed into blocks by an adaptive algorithm based on image content to obtain a plurality of image blocks, including:

[0110] Preprocessing the real optical image, wherein the preprocessing includes denoising and contrast enhancement;

[0111] Extracting feature information of the preprocessed real optical image based on image processing technology, wherein the feature information includes at least spatial features, texture features and color features;

[0112] Calculating the similarity between different regions of the real optical image according to the feature information;

[0113] Based on the calculation result of the similarity, the real optical image is divided into a plurality of image blocks.

[0114] Specifically, the spatial features of objects in the image can be described by edge detection and corner detection, the texture features in the image can be extracted by gray-level co-occurrence matrix and local binary pattern, and the color features of the image can be described by color histogram. Next, the cosine value of the angle between the two feature vectors is calculated. The larger the value, the higher the similarity. Using algorithms such as K-means and hierarchical clustering, the image is divided into multiple image blocks according to the similarity of the feature vectors.

[0115] Preferably, in one embodiment of the present invention, extracting each pixel point in each of the image blocks as a graph node, calculating the weight of the edge according to the adjacency relationship and feature similarity between any two of the pixel points, and constructing graph structure data according to the graph nodes and the weights of the edges, comprises:

[0116] In each image block, each pixel point is characterized as a graph node, and the graph node at least includes the spatial feature, texture feature and color feature of the pixel point;

[0117] Determine the adjacency relationship between any two of the graph nodes based on the spatial features, calculate the feature similarity between any two of the pixel points based on the texture features and the color features, and calculate the edge weight corresponding to each of the graph nodes based on the adjacency relationship and the feature similarity;

[0118] Graph structure data is constructed according to the graph nodes and the edge weights.

[0119] Specifically, in each image block, each pixel is regarded as a graph node. Each graph node contains not only the location information of the pixel (i.e., spatial features), but also its texture features and color features. The edge weight is calculated based on the adjacency relationship and feature similarity between graph nodes. The edge weight can reflect the connection strength or similarity between two graph nodes.

[0120] Graph structure data consists of graph nodes and edges. Each graph node contains the spatial features, texture features, and color features of the pixel. Each edge connects two graph nodes and has a weight value. Based on the graph nodes and edge weights, a graph structure data can be constructed. This graph can be used for subsequent graph processing and analysis tasks, such as image segmentation, image recognition, etc.

[0121] Preferably, in one embodiment of the present invention, the feature encoding of the graph structure data based on the graph attention network to obtain the encoded feature vector, and decoding the feature vector to obtain the decoded feature map includes:

[0122] Inputting the constructed graph structure data into the graph attention network, iteratively updating the nodes in the graph structure data through the multi-layer attention mechanism of the graph attention network, and obtaining the feature vector representation of each node;

[0123] The feature vector of each node is converted by a decoder to obtain a decoded feature map.

[0124] Specifically, in the graph attention network, initialize the weights and bias parameters of all layers. These parameters will be optimized by the back-propagation algorithm during the training process. Input the graph structure data to the input layer of the graph attention network. This usually includes a node feature matrix (each node corresponds to a row and each feature corresponds to a column) and an adjacency matrix (or edge list, which represents the connection relationship between nodes).

[0125] For each node, calculate its attention coefficient with its neighboring nodes. This is usually achieved by calculating the similarity between node features (such as dot product, cosine similarity, etc.) and applying a softmax function. According to the attention coefficient, the features of the neighboring nodes are weighted summed to obtain the updated feature vector of the node. Repeat this process until the feature vectors of all nodes are updated.

[0126] The updated node feature vector of the first layer is used as the input of the next layer, and the above attention mechanism process is repeated. Each layer may contain nonlinear activation functions (such as ReLU) and / or normalization layers (such as Layer Normalization) to enhance the expressiveness and stability of the network. Through multiple layers of iteration, the node feature vector will gradually fuse global information to form a richer representation. In the last layer, the final feature vector representation of each node is output. These feature vectors contain the information of the node itself as well as the global information fused through the multi-layer attention mechanism.

[0127] Design a decoder network to convert node feature vectors into decoded feature maps. Input the feature vector of each node into the decoder. The decoder outputs the feature value of the corresponding position based on the input feature vector. These feature values ​​can form a new feature map, in which the value of each pixel is determined by the feature vector of the corresponding node. If needed, post-process the decoded feature map, such as normalization, smoothing, etc., to improve its visual effect or meet specific application requirements. Finally, output the decoded feature map, which contains rich feature information of each pixel in the original image and may have higher resolution or finer details.

[0128] Furthermore, after obtaining the decoded feature map, it also includes:

[0129] Acquire a spectral image of the target degradation film material scanned by a spectral detection device, and perform feature extraction on the spectral image to obtain a spectral image feature graph;

[0130] The feature map and the spectral image feature map are subjected to multimodal image fusion to obtain fusion features.

[0131] Wherein, performing multimodal image fusion on the feature map and the spectral image feature map includes:

[0132] Establishing a correspondence between the feature map and the spectral image feature map by a feature point matching algorithm;

[0133] Spatially aligning the feature map and the spectral image feature map according to the relationship;

[0134] Through the improved feature fusion algorithm, a fused feature map containing optical image features and spectral image features is generated.

[0135] Specifically, the multimodal image fusion of optical image feature maps and spectral image feature maps based on feature-level fusion methods is a complex and delicate process, which requires comprehensive consideration of multiple links such as feature extraction, matching, alignment, fusion, and post-processing. By reasonably selecting and combining these steps and methods, a fused feature map with higher information content and better representation ability can be generated, providing strong support for practical applications.

[0136] Preferably, in one embodiment of the present invention, the characteristic graph is input into each of the identification models to obtain the matching degree between the target degradation film material and each type of degradation film material, and the type of the target degradation film material is identified according to the matching degree, including:

[0137] Inputting the characteristic graph of the target degradation film material into each of the identification models respectively, and obtaining the matching degree between the target degradation film material and the corresponding degradation mold material type, wherein the matching degree reflects the similarity between the target degradation film material and the corresponding degradation mold material;

[0138] The matching degrees outputted by the various recognition models are compared, and the type of degradation film material corresponding to the model with the highest matching degree is selected as the type of the target degradation film material.

[0139] Specifically, a plurality of recognition models are collected and trained, the model being capable of receiving a feature map as input and outputting a degree of matching with a target degradation film material type.

[0140] The characteristic graph of the target degradation film material is input into each prepared recognition model respectively. Each recognition model will receive the characteristic graph as input and calculate the matching degree with the target degradation film material type based on its internal algorithm and training data. The matching degree is usually a value between 0 and 1, indicating the similarity between the target degradation film material and the corresponding degradation film material type.

[0141] As needed, post-process the selected degradation membrane material type, such as further verifying its accuracy, recording the results, etc.

[0142] Another embodiment of the present application provides a PBAT degradation film material type identification system based on image processing technology. For details, please refer to Figure 3 , Figure 3 The structure diagram of the PBAT degradation film material type identification system based on image processing technology provided in one embodiment of the present application is shown. It includes: a compensation module 11, a division module 12, an extraction module 13, an encoding module 14, a construction module 15, and an identification module 16, wherein

[0143] A compensation module 11 is used to perform shadow compensation on an initial optical image of a target degradation film material taken by an optical imaging device to obtain a real optical image of the target degradation film material, wherein the shadow compensation is designed to eliminate the influence of the missing depth value of a transparent object on the recognition accuracy of the image of the target degradation film material;

[0144] A dividing module 12, configured to divide the real optical image into blocks by using an adaptive algorithm based on image content to obtain a plurality of image blocks;

[0145] An extraction module 13, used to extract each pixel point in each of the image blocks as a graph node, and calculate the weight of the edge according to the adjacency relationship and feature similarity between any two of the pixel points, and construct graph structure data according to the graph nodes and the weights of the edges;

[0146] An encoding module 14 is used to perform feature encoding on the graph structure data based on a graph attention network to obtain an encoded feature vector, and decode the feature vector to obtain a decoded feature graph;

[0147] A construction module 15 is used to construct a corresponding number of recognition models based on the type of degradation film material, wherein each recognition model corresponds to one type of degradation film material;

[0148] The identification module 16 is used to input the characteristic graph into each of the identification models respectively, obtain the matching degree between the target degradation film material and each type of degradation film material, and identify the type of the target degradation film material according to the matching degree.

[0149] Compared with the prior art, the beneficial effects of the embodiments of the present application are at least one of the following:

[0150] (1) This application eliminates the impact of the lack of depth value of transparent objects on image recognition accuracy through shadow compensation technology. As a transparent or translucent object, PBAT degradation film material is easily affected by lighting conditions during the imaging process, resulting in shadows or reflections in the image. Shadow compensation technology can effectively reduce these interference factors, making the image more realistically reflect the characteristics of the material, thereby improving the accuracy of recognition.

[0151] (2) This application uses an adaptive algorithm based on image content to process real optical images in blocks, which can be flexibly divided according to the feature differences of different regions in the image, so that each image block can better reflect the local characteristics of the material. This processing method helps to capture more detailed information and further improve the accuracy of recognition.

[0152] (3) This application uses a graph attention network to encode and decode features of graph structure data, which can automatically learn the complex relationships between pixels in the image and highlight key features. This method can more accurately capture the unique texture, color and other features of PBAT degradation film materials, thereby improving recognition accuracy.

[0153] (4) Compared with manual classification methods, this application can significantly reduce the classification error rate. This helps to reduce additional costs such as rework and scrapping of finished products caused by inaccurate classification. Accurate identification methods help ensure that the quality of PBAT degradation film materials meets standard requirements.

[0154] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for identifying the type of PBAT degradation film material based on image processing technology, characterized in that: The following steps are involved: Performing shadow compensation on an initial optical image of a target degradation film material taken by an optical imaging device to obtain a true optical image of the target degradation film material, wherein the shadow compensation is designed to eliminate the influence of the missing depth value of a transparent object on the recognition accuracy of the image of the target degradation film material; The real optical image is processed by a block-by-block process using an adaptive algorithm based on image content to obtain a plurality of image blocks; Taking each pixel point extracted from each of the image blocks as a graph node, calculating the weight of an edge according to the adjacency relationship and feature similarity between any two of the pixel points, and constructing graph structure data according to the graph nodes and the weights of the edges; Based on the graph attention network, feature encoding is performed on the graph structure data to obtain an encoded feature vector, and the feature vector is decoded to obtain a decoded feature graph; Constructing a corresponding number of recognition models based on the type of degradation film material, wherein each recognition model corresponds to one type of degradation film material; The characteristic graphs are input into the respective identification models to obtain the matching degree between the target degradation film material and the types of the degradation film materials, and the type of the target degradation film material is identified according to the matching degree.

2. The PBAT degradation film material type identification method based on image processing technology according to claim 1 is characterized in that: The shadow compensation is performed on the initial optical image of the target degradation film material taken by the optical imaging device to obtain the real optical image of the target degradation film material, including: Inputting the initial optical image into a pre-trained shadow detection model to perform shadow detection, thereby obtaining a shadow region of the initial optical image, wherein the shadow detection model is designed to correct a prediction result of the shadow detection model through real-time shadow error detection; Processing the initial optical image according to the shadow region and the shadow removal algorithm to obtain a shadow-free image; Extracting the depth value of the initial optical image to obtain the depth information of the target degradation film material; The shadow-free image and the depth information are processed based on a stereo matching algorithm to obtain a true optical image of the target degradation film material.

3. The PBAT degradation film material type identification method based on image processing technology according to claim 2 is characterized in that: The processing of the initial optical image according to the shadow area and the shadow removal algorithm to obtain a shadow-free image includes: Converting the initial optical image into the HSV color space to obtain an HSV image; Distinguish the shadow area and the non-shadow area by the brightness value of the HSV image; The brightness value of the shadow area is corrected based on the brightness value of the non-shadow area, which is expressed as: ； in, represents the corrected HSV image, represents the HSV image before correction, and is the adjustment factor, is the brightness threshold; The corrected HSV image is converted into the RGB color space to obtain the shadow-free image.

4. The PBAT degradation film material type identification method based on image processing technology according to claim 2 is characterized in that: The extracting the depth value of the initial optical image to obtain the depth information of the target degradation film material includes: Dividing the space where the initial optical image is located into a voxel grid; Simulating rays emitted by an imaging device based on a ray tracing algorithm, and calculating intersection points of each ray with the voxel grid according to an intersection detection algorithm; The characteristic information of the intersection is obtained, and the characteristic information is input into a pre-trained depth prediction model to obtain the depth information of the target degradation film material.

5. The PBAT degradation film material type identification method based on image processing technology according to claim 4 is characterized in that: Before inputting the feature information into the pre-trained depth prediction model, the method further includes: Collecting an RGB-D image data set of the degradation film material, wherein the RGB-D image data set includes depth information of the degradation film material; Dividing the space where the RGB-D image data set is located into a voxel grid, extracting target features from the RGB-D image data set, mapping the target features to the voxel space, generating a voxel feature vector for each voxel, and obtaining a voxel feature vector set; A basic depth prediction model based on a deep convolutional neural network is constructed, and the basic depth prediction model is trained based on the voxel feature vector set to obtain a depth prediction model.

6. The PBAT degradation film material type identification method based on image processing technology according to claim 1 is characterized in that: The real optical image is processed into blocks by an adaptive algorithm based on image content to obtain a plurality of image blocks, including: Preprocessing the real optical image, wherein the preprocessing includes denoising and contrast enhancement; Extracting feature information of the preprocessed real optical image based on image processing technology, wherein the feature information includes at least spatial features, texture features and color features; Calculating the similarity between different regions of the real optical image according to the feature information; Based on the calculation result of the similarity, the real optical image is divided into a plurality of image blocks.

7. The PBAT degradation film material type identification method based on image processing technology according to claim 1 is characterized in that: The method of taking each pixel point extracted from each of the image blocks as a graph node, calculating the weight of an edge according to the adjacency relationship and feature similarity between any two of the pixel points, and constructing graph structure data according to the graph nodes and the weights of the edges includes: In each image block, each pixel point is characterized as a graph node, and the graph node at least includes the spatial feature, texture feature and color feature of the pixel point; Determine the adjacency relationship between any two of the graph nodes based on the spatial features, calculate the feature similarity between any two of the pixel points based on the texture features and the color features, and calculate the edge weight corresponding to each of the graph nodes based on the adjacency relationship and the feature similarity; Graph structure data is constructed according to the graph nodes and the edge weights.

8. The method for identifying the type of PBAT degradation film material based on image processing technology according to claim 1 is characterized in that: The step of encoding the graph structure data based on the graph attention network to obtain an encoded feature vector, and decoding the feature vector to obtain a decoded feature graph includes: Inputting the constructed graph structure data into the graph attention network, iteratively updating the nodes in the graph structure data through the multi-layer attention mechanism of the graph attention network, and obtaining the feature vector representation of each node; The feature vector of each node is converted by a decoder to obtain a decoded feature map.

9. The method for identifying the type of PBAT degradation film material based on image processing technology according to claim 1, characterized in that: The step of inputting the characteristic graph into each of the identification models to obtain the matching degree between the target degradation film material and each type of degradation film material, and identifying the type of the target degradation film material according to the matching degree comprises: Inputting the characteristic graph of the target degradation film material into each of the identification models respectively, and obtaining the matching degree between the target degradation film material and the corresponding degradation mold material type, wherein the matching degree reflects the similarity between the target degradation film material and the corresponding degradation mold material; The matching degrees outputted by the various recognition models are compared, and the type of degradation film material corresponding to the model with the highest matching degree is selected as the type of the target degradation film material.

10. A PBAT degradation film material type identification system based on image processing technology, characterized in that: Includes the following modules: A compensation module, used for performing shadow compensation on an initial optical image of a target degradation film material taken by an optical imaging device to obtain a real optical image of the target degradation film material, wherein the shadow compensation is designed to eliminate the influence of the missing depth value of a transparent object on the recognition accuracy of the image of the target degradation film material; A partitioning module, configured to perform a partitioning process on the real optical image by using an adaptive algorithm based on image content to obtain a plurality of image blocks; An extraction module, used to take each pixel point extracted from each of the image blocks as a graph node, calculate the weight of an edge according to the adjacency relationship and feature similarity between any two of the pixel points, and construct graph structure data according to the graph nodes and the weights of the edges; An encoding module, used for performing feature encoding on the graph structure data based on a graph attention network to obtain an encoded feature vector, and decoding the feature vector to obtain a decoded feature graph; A construction module, used to construct a corresponding number of recognition models based on the type of degradation film material, wherein each of the recognition models corresponds to one type of degradation film material; The identification module is used to input the characteristic graph into each of the identification models respectively, obtain the matching degree between the target degradation film material and each type of degradation film material, and identify the type of the target degradation film material according to the matching degree.

Images