A convolutional neural network-based method and system for identifying peptide drug membrane filtration contamination

Abstract

The present invention relates to the field of target recognition technology, and specifically discloses a method and system for identifying polypeptide drug membrane filtration pollution using a convolutional neural network, wherein the method comprises the following steps: obtaining an image to be identified and a corresponding image sequence; generating a corrected image sequence including the corrected image to be identified; inputting the corrected image to be identified in the corrected image sequence to obtain a high-resolution feature map and a low-resolution feature map; analyzing local area characteristics of the corrected image to be identified to generate a weight parameter group; weightedly fusing the high-resolution feature map and the low-resolution feature map according to the weight parameter group to generate a fused feature map; performing target detection based on the fused feature map to generate a pollution identification result; the method effectively addresses the problems of slight shaking and detail loss under complex imaging conditions, thereby achieving accurate identification of slight pollution targets, having better robustness, being able to effectively identify slight pollution targets, and reducing false alarm and omission rates.

Description

Technical Field

[0001] The present application relates to the field of target recognition technology, and more specifically, to a method and system for identifying polypeptide drug membrane filtration contamination using a convolutional neural network. Background Art

[0002] In peptide drug production, membrane filtration is a key purification step, used to remove impurities to ensure product quality and drug safety. During production, membrane filtration equipment operates continuously, pumping the purified solution through a membrane with a specific pore size, and the filtrate is used for subsequent processing.

[0003] During filtration, there is a risk of contamination on the filter membrane surface or in the solution, and existing technologies mostly use methods based on image analysis for monitoring. Image acquisition units, including high-resolution cameras and light sources, are installed at key locations on the equipment to periodically or continuously acquire image data and transmit it to the processing unit. An image recognition program based on a convolutional neural network (CNN) analyzes the data to determine contamination, identify the type, determine the location, estimate the size, and record the information. If necessary, it can trigger an alarm or pause the process. This method improves detection efficiency, consistency, and objectivity. However, actual imaging conditions are complex and dynamically changing. CNN models based on idealized training data are difficult to adapt to, and it is difficult to identify tiny contamination targets. False positives or omissions are prone to occur, and they cannot meet the strict requirements of production for product purity control.

[0004] There is currently no effective technical solution to the above problems. Summary of the Invention

[0005] The purpose of this application is to provide a convolutional neural network method and system for identifying polypeptide drug membrane filtration contamination to address the problems of slight shaking and detail loss under complex imaging conditions, so as to accurately identify tiny contamination targets and reduce false alarm and missed alarm rates.

[0006] In a first aspect, the present application provides a method for identifying polypeptide drug membrane filtration contamination using a convolutional neural network, the method comprising the following steps:

[0007] S1. Obtain the image to be identified and the corresponding image sequence on the surface of the filter membrane;

[0008] S2. performing position correction on the image sequence to generate a corrected image sequence including corrected images to be recognized;

[0009] S3, inputting the corrected image to be recognized in the corrected image sequence into a preset high-resolution processing path and a preset low-resolution processing path to obtain a high-resolution feature map and a low-resolution feature map;

[0010] S4, analyzing the local area characteristics of the corrected image to be identified to generate a weight parameter group;

[0011] S5. Weightedly fuse the high-resolution feature map and the low-resolution feature map according to the weight parameter group to generate a fused feature map;

[0012] S6. Perform target detection based on the fused feature map to generate a pollution identification result.

[0013] The core innovation of the convolutional neural network method for identifying polypeptide drug membrane filtration pollution in this application lies in the combination of high- and low-resolution dual-path processing and weighted fusion based on local area characteristics, thereby generating a fusion feature map for target detection, effectively dealing with the problems of slight shaking and detail loss under complex imaging conditions, thereby achieving accurate identification of tiny pollution targets, having better robustness, being able to effectively identify tiny pollution targets, and reducing false alarm and missed alarm rates.

[0014] In the method for identifying polypeptide drug membrane filtration contamination using a convolutional neural network, the local region characteristics include local contrast, local texture complexity, and local motion degree, and step S4 includes:

[0015] S41, calculating the gradient magnitude and gradient direction of each pixel in the corrected image to be identified;

[0016] S42, dividing the corrected image to be recognized into a plurality of image blocks according to the gradient magnitude and the gradient direction, and calculating the local contrast, local texture complexity, and local motion degree of each image block;

[0017] S43, calculating initial weight parameters corresponding to the high-resolution processing path and the low-resolution processing path for each image block using a preset nonlinear mapping model based on the local contrast, local texture complexity, and local motion degree of each image block;

[0018] S44: Normalize the initial weight parameters to obtain a final weight parameter group.

[0019] Through the above technical solution, the present application can generate a more refined and discriminative weight parameter group based on specific characteristics such as local contrast, local texture complexity, and local motion degree of the corrected image to be identified.

[0020] The method for identifying polypeptide drug membrane filtration contamination using a convolutional neural network, wherein step S43 includes:

[0021] S431, for each image block, constructing a multidimensional feature vector including local contrast, local texture complexity, and local motion degree;

[0022] S432. Input the multidimensional feature vector into a pre-trained nonlinear mapping model to output initial weight parameters corresponding to a high-resolution processing path and a low-resolution processing path for each image block, wherein the initial weight parameters of the high-resolution processing path are positively correlated with the local contrast and the local texture complexity, and negatively correlated with the local motion degree.

[0023] In the method for identifying polypeptide drug membrane filtration contamination using a convolutional neural network, the step of calculating the local contrast, local texture complexity, and local motion degree of each image block includes:

[0024] S423, calculating the standard deviation of the grayscale values of all pixels in each image block, and dividing the standard deviation by the mean grayscale values of all pixels in the image block to obtain a normalized local contrast;

[0025] S424, calculating the histogram of the gradient direction of all pixels in each image block, counting the number of pixels in each gradient direction, and calculating the entropy value of the histogram as the local texture complexity;

[0026] S425 , calculating and obtaining the optical flow vector between each image block and the corresponding image block of the previous frame, and calculating the average amplitude of the optical flow vector to obtain the degree of local motion.

[0027] In the method for identifying polypeptide drug membrane filtration contamination using a convolutional neural network, the step of dividing the corrected image to be identified into multiple image blocks according to the gradient amplitude and the gradient direction includes:

[0028] S421, setting the weight of the pixel point according to the gradient amplitude, the larger the gradient amplitude, the higher the weight;

[0029] S422: Using a weighted K-means clustering algorithm, with the position coordinates of the pixel points as features and the weights of the pixel points as clustering weights, the corrected image to be identified is divided into multiple image blocks.

[0030] The method for identifying polypeptide drug membrane filtration contamination using a convolutional neural network, wherein step S41 comprises:

[0031] S411, using the Sobel operator to calculate the gradient components of the corrected image to be identified in the horizontal and vertical directions respectively, to obtain a horizontal gradient map and a vertical gradient map;

[0032] S412, calculating the gradient amplitude of each pixel according to the horizontal gradient map and the vertical gradient map, and processing the gradient amplitude using a non-maximum suppression algorithm to obtain a refined gradient amplitude map;

[0033] S413 , calculating the gradient direction of each pixel according to the horizontal gradient map and the vertical gradient map to obtain a discretized gradient direction map.

[0034] The method for identifying polypeptide drug membrane filtration contamination using a convolutional neural network, wherein step S5 comprises:

[0035] S51, for the high-resolution feature map and the low-resolution feature map, respectively constructing a weight parameter matrix corresponding to the corrected image according to the weight parameter group;

[0036] S52, extracting a high-resolution weight parameter of an image block corresponding to each pixel in the high-resolution feature map according to a weight parameter matrix corresponding to the high-resolution feature map, and adjusting a feature vector corresponding to the pixel in the high-resolution feature map according to the high-resolution weight parameter to obtain an adjusted high-resolution feature map;

[0037] S53, extracting a low-resolution weight parameter of an image block corresponding to each pixel in the low-resolution feature map according to a weight parameter matrix corresponding to the low-resolution feature map, and adjusting a feature vector corresponding to the pixel in the low-resolution feature map according to the low-resolution weight parameter to obtain an adjusted low-resolution feature map;

[0038] S54: Concatenate the adjusted high-resolution feature map and the adjusted low-resolution feature map in the channel dimension to obtain a fused feature map.

[0039] The method for identifying polypeptide drug membrane filtration contamination using a convolutional neural network, wherein step S3 comprises:

[0040] S31, preprocessing the corrected image to be recognized to obtain a preprocessed image;

[0041] S32, inputting the preprocessed image into a preset high-resolution convolutional neural network to extract a high-resolution feature map, wherein the high-resolution convolutional neural network includes multiple convolutional layers and pooling layers;

[0042] S33. Input the preprocessed image into a preset low-resolution convolutional neural network to extract a low-resolution feature map. The low-resolution convolutional neural network includes multiple convolution layers, pooling layers and downsampling layers.

[0043] The method for identifying polypeptide drug membrane filtration contamination using a convolutional neural network, wherein step S6 comprises:

[0044] S61, performing a convolution operation on the fused feature map and each convolution kernel in a preset convolution kernel set to obtain a plurality of response maps corresponding to pollution types, wherein each convolution kernel in the convolution kernel set performs response enhancement for different types of pollution targets;

[0045] S62. For each response map, a non-maximum suppression algorithm is used for post-processing to screen candidate boxes of contaminated targets with high confidence. The pollution identification result is output based on the position, size, and response intensity of the candidate boxes and the mapping relationship between pollution type and response intensity.

[0046] In a second aspect, the present application also provides a convolutional neural network polypeptide drug membrane filtration contamination identification system, the system comprising:

[0047] An acquisition module, used to acquire the image to be identified and the corresponding image sequence on the surface of the filter membrane;

[0048] a correction module, configured to perform position correction on the image sequence to generate a corrected image sequence containing corrected images to be recognized;

[0049] a processing module, configured to input the corrected image to be identified in the corrected image sequence into a preset high-resolution processing path and a preset low-resolution processing path to obtain a high-resolution feature map and a low-resolution feature map;

[0050] A weight module, used for analyzing the local area characteristics of the corrected image to be identified to generate a weight parameter group;

[0051] A fusion module, configured to perform weighted fusion of the high-resolution feature map and the low-resolution feature map according to the weight parameter group to generate a fused feature map;

[0052] The recognition module is used to perform target detection based on the fused feature map to generate a pollution recognition result.

[0053] The convolutional neural network polypeptide drug membrane filtration pollution identification system of the present application combines high- and low-resolution dual-path processing and weighted fusion based on local area characteristics to generate a fusion feature map for target detection, effectively dealing with the problems of slight shaking and detail loss under complex imaging conditions, thereby achieving accurate identification of tiny contaminated targets.

[0054] As can be seen from the above, the present application provides a method and system for identifying polypeptide drug membrane filtration pollution using a convolutional neural network, wherein the method of the present application solves the problem of poor robustness and easy false positives and missed reports when existing convolutional neural network methods identify tiny pollution targets of polypeptide drug membrane filtration under complex dynamic imaging conditions. It eliminates the interference of image shaking on recognition and improves the stability of processing by performing position correction on the image sequence. By processing high-resolution and low-resolution paths in parallel, it captures the detail information important to tiny targets and low-frequency features that are robust to environmental changes respectively. By analyzing local area characteristics and performing weighted fusion, the feature representation can be optimized according to the actual conditions of different areas of the image, thereby enhancing the sensitivity to tiny targets and improving the adaptability to complex imaging conditions. Target detection based on the optimized fusion feature map can more accurately distinguish between real pollution targets and background interference, thereby improving the recognition accuracy and robustness of tiny pollution targets and reducing the false positive and missed reports. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a flow chart of a method for identifying polypeptide drug membrane filtration contamination using a convolutional neural network provided in an embodiment of the present application.

[0056] Figure 2 A schematic diagram of the structure of a convolutional neural network polypeptide drug membrane filtration contamination identification system provided in an embodiment of the present application.

[0057] Reference numerals: 201, acquisition module; 202, correction module; 203, processing module; 204, weight module; 205, fusion module; 206, recognition module. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work fall within the scope of protection of the present application.

[0059] It should be noted that similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and should not be understood as indicating or implying relative importance.

[0060] First, please refer to Figure 1 Some embodiments of the present application provide a method for identifying polypeptide drug membrane filtration contamination using a convolutional neural network, the method comprising the following steps:

[0061] S1. Obtain the image to be identified and the corresponding image sequence on the surface of the filter membrane;

[0062] S2. Perform position correction on the image sequence to generate a corrected image sequence including the corrected images to be recognized;

[0063] S3, inputting the corrected image to be recognized in the corrected image sequence into a preset high-resolution processing path and a preset low-resolution processing path to obtain a high-resolution feature map and a low-resolution feature map;

[0064] S4, analyzing the local area characteristics of the corrected image to be identified to generate a weight parameter group;

[0065] S5. Weightedly fuse the high-resolution feature map and the low-resolution feature map according to the weight parameter group to generate a fused feature map;

[0066] S6. Perform target detection based on the fused feature map to generate pollution identification results.

[0067] Specifically, position correction involves adjusting the spatial positions of individual frames in an image sequence so that they align with one another. This can be achieved by analyzing the pixel offsets between frames and applying a transformation. Position correction compensates for minor shakes or drifts during image acquisition, ensuring spatial alignment between frames in the sequence and ensuring spatial consistency within the sequence, providing stable input for subsequent feature extraction and analysis.

[0068] More specifically, the high-resolution and low-resolution processing paths involve feeding the image into two distinct computational processes. The high-resolution processing path preserves high-frequency image details, while the low-resolution processing path extracts robust low-frequency features. Thus, the high-resolution processing path extracts features that preserve fine image details, while the low-resolution processing path extracts features that are adaptable to image changes. This yields feature representations that are both detail-sensitive and adaptable to changes. These two paths can be implemented using different convolutional neural network architectures.

[0069] More specifically, local region feature analysis involves calculating the properties of small regions within an image. These properties can include contrast, texture, and motion. This can be achieved by calculating statistics within the region. This provides information about the content and properties of different regions in the image.

[0070] More specifically, the weight parameter group refers to a set of values generated based on the results of local region feature analysis. These values are used to guide the subsequent feature fusion process.

[0071] More specifically, weighted fusion involves combining high-resolution and low-resolution feature maps based on a set of weight parameters. This can be achieved by applying weights to the feature maps and then concatenating or summing them. This adaptively integrates different types of features based on the characteristics of local image regions, generating a more comprehensive representation.

[0072] More specifically, object detection involves identifying, locating, and determining the size and type of contaminants within a fused feature map. This is achieved by convolving the feature map with a convolution kernel and performing post-processing. This outputs the final contamination identification results, including the location, size, and type of tiny contaminants.

[0073] More specifically, the method of the present application solves the problem of poor robustness and easy false alarms and missed detections of existing convolutional neural network methods when identifying tiny pollution targets filtered by polypeptide drug membranes under complex dynamic imaging conditions. It eliminates the interference of image shaking on recognition and improves the stability of processing by performing position correction on the image sequence. By processing high-resolution and low-resolution paths in parallel, it captures the detail information important for tiny targets and low-frequency features robust to environmental changes respectively. By analyzing local area characteristics and performing weighted fusion, the feature representation can be optimized according to the actual conditions of different areas of the image, enhancing the sensitivity to tiny targets and improving the adaptability to complex imaging conditions. Target detection based on the optimized fusion feature map can more accurately distinguish between real pollution targets and background interference, thereby improving the recognition accuracy and robustness of tiny pollution targets and reducing the false alarm and missed detection rates.

[0074] Therefore, the core innovation of the polypeptide drug membrane filtration pollution identification method of the convolutional neural network in the embodiment of the present application lies in the combination of high- and low-resolution dual-path processing and weighted fusion based on local area characteristics, thereby generating a fusion feature map for target detection, effectively dealing with the problems of slight shaking and detail loss under complex imaging conditions, thereby achieving accurate identification of tiny pollution targets, having better robustness, being able to effectively identify tiny pollution targets, and reducing false alarm and missed alarm rates.

[0075] In some preferred embodiments, the local region characteristics include local contrast, local texture complexity, and local motion degree, and step S4 includes:

[0076] S41, calculating the gradient magnitude and gradient direction of each pixel in the corrected image to be identified;

[0077] S42, dividing the corrected image to be identified into multiple image blocks according to the gradient amplitude and gradient direction, and calculating the local contrast, local texture complexity, and local motion degree of each image block;

[0078] S43, calculating initial weight parameters corresponding to the high-resolution processing path and the low-resolution processing path for each image block using a preset nonlinear mapping model based on the local contrast, local texture complexity, and local motion degree of each image block;

[0079] S44: Normalize the initial weight parameters to obtain a final weight parameter group.

[0080] Specifically, in step S41 , the gradient amplitude represents the severity of the pixel grayscale change, which is related to the edge, details and contrast of the image; the gradient direction indicates the direction of the grayscale change, which is related to the texture structure of the image.

[0081] More specifically, in step S42, image block segmentation based on gradient information helps group regions with similar structures or variation characteristics into the same block. Within each image block, local contrast, local texture complexity, and local motion are calculated. These specific local characteristics quantify the properties of the image block from different dimensions, laying the foundation for the subsequent generation of targeted weight parameters.

[0082] More specifically, in step S43, the nonlinear mapping model can be constructed using a deep neural network to fit the nonlinear relationship between the multidimensional feature vector and the weights of the high- and low-resolution processing paths. Weight parameters are calculated separately for the high- and low-resolution paths because different local characteristics have different requirements for features of different resolutions: high-contrast or complex textured areas may require higher weights for high-resolution features that preserve detail, while moving or flat areas may rely more on robust low-resolution features. Calculating weights based on specific local characteristics allows the weight parameters to more accurately reflect each local area's preference for high- and low-resolution features.

[0083] More specifically, in step S44, the normalization process ensures that the weight parameters are within a reasonable range, facilitates the subsequent weighted fusion operation, and ensures the stability of the fusion process.

[0084] More specifically, through the above technical solution, the present application can generate a more refined and discriminative weight parameter group based on the specific characteristics of the corrected image to be identified, such as the local contrast, local texture complexity, and local motion degree. As a result, in the subsequent feature fusion step, the contribution ratio of the high-resolution feature map and the low-resolution feature map can be more accurately adjusted according to these weight parameters, so that the fused feature map can better retain the detailed information of the small target while suppressing background interference and motion artifacts. This helps to improve the accuracy of subsequent target detection, so that small contaminated targets can be effectively distinguished and identified, reducing false positives and missed positives.

[0085] In some preferred embodiments, step S43 includes:

[0086] S431, for each image block, constructing a multidimensional feature vector including local contrast, local texture complexity, and local motion degree;

[0087] S432. Input the multidimensional feature vector into a pre-trained nonlinear mapping model to output initial weight parameters corresponding to the high-resolution processing path and the low-resolution processing path for each image block. The initial weight parameters of the high-resolution processing path are positively correlated with the local contrast and local texture complexity, and negatively correlated with the degree of local motion.

[0088] Specifically, the above processing method aims to provide a more accurate and robust way to calculate the weight parameters of high- and low-resolution feature fusion, so as to overcome the shortcomings of simple preset nonlinear functions in processing complex local area characteristics and weight relationships.

[0089] More specifically, step S431 quantifies the local region characteristics of the image block and integrates them into a unified input form. Local contrast, local texture complexity, and local motion describe the visual and dynamic characteristics of the image block from different dimensions. Combining them into a vector can more comprehensively represent the local information of the image block, providing a rich and structured input for subsequent weight calculation.

[0090] More specifically, step S432 inputs the local region feature vector into a pre-trained model. The model can directly output the weights that should be assigned to the image block on the high-resolution path and the low-resolution path based on the complex mapping relationship it has learned. The nonlinear mapping model is composed of a deep neural network, which fits the nonlinear relationship between the multidimensional feature vector and the weights of the high- and low-resolution processing paths. The weight of the high-resolution path is positively correlated with the local contrast and local texture complexity, and negatively correlated with the degree of local motion. This is because the high-resolution path focuses on preserving details, while high-contrast and complex texture areas usually contain rich detail information, and motion may cause details to be blurred. The weight of the low-resolution path is negatively correlated with the local contrast and local texture complexity, and positively correlated with the degree of local motion. This is because the low-resolution path focuses on extracting robust features. These features are insensitive to details but may be more stable in areas with less obvious motion or details. By clarifying this correlation, a guiding direction is provided for model training.

[0091] More specifically, the nonlinear mapping model designed by the deep neural network has powerful nonlinear fitting capabilities, capable of learning and representing complex relationships between local characteristics and weights that are difficult to describe with simple functions, thus overcoming the limitations of simple pre-set nonlinear functions. Compared to using simple pre-set functions to calculate weights, the pre-trained deep neural network can learn and adapt to more complex relationships between local characteristics and weights, thereby generating more precise weight parameters, significantly improving the effectiveness of feature fusion and the final pollution identification performance. Its advantages are particularly evident when dealing with complex and changing real-world imaging conditions and identifying small targets.

[0092] More specifically, by using a deep neural network for nonlinear mapping, the calculated initial weight parameters can more accurately reflect the requirements of the local characteristics of the current image block for the fusion of high- and low-resolution features. These more accurate initial weight parameters are normalized and used to weightedly fuse the high- and low-resolution feature maps. This improved weight calculation method makes feature fusion more effective, better preserving important details while suppressing the effects of noise and motion blur. This is especially true for the recognition of small contaminated objects, where detailed information is crucial. Accurate weight allocation can significantly improve recognition accuracy and robustness.

[0093] In some preferred embodiments, the step of calculating the local contrast, local texture complexity, and local motion degree of each image block includes:

[0094] S423, calculating the standard deviation of the grayscale values of all pixels in each image block, and dividing the standard deviation by the mean grayscale values of all pixels in the image block to obtain a normalized local contrast;

[0095] S424, calculating the histogram of the gradient direction of all pixels in each image block, counting the number of pixels in each gradient direction, and calculating the entropy value of the histogram as the local texture complexity;

[0096] S425 , calculating and obtaining the optical flow vector between each image block and the corresponding image block of the previous frame, and calculating the average amplitude of the optical flow vector to obtain the degree of local motion.

[0097] Specifically, in the calculation process of local contrast, the standard deviation measures the dispersion of pixel grayscale values, reflecting the local contrast, and is divided by the mean for normalization to eliminate the influence of the overall brightness of the image block on the contrast measurement and make the result stable.

[0098] More specifically, when calculating local texture complexity, the gradient direction histogram describes the distribution of edge and texture directions within an image patch. Entropy measures the uncertainty of information and, in this case, the uniformity of the gradient direction distribution. Higher entropy values indicate a more uniform distribution of gradient directions and a more complex texture.

[0099] More specifically, when calculating the degree of local motion, the optical flow vector reflects the motion of pixels between consecutive frames, and its magnitude indicates the speed of motion. Calculating the average magnitude can quantify the overall intensity of motion in a local area. For example, an optical flow algorithm such as the Lucas-Kanade algorithm can be used to calculate the optical flow vectors of pixels within an image block, and then the average magnitude of these vectors can be calculated.

[0100] More specifically, the quantification results of local contrast, local texture complexity, and local motion obtained through these specific methods are accurate and stable. These quantification results are used to subsequently generate weight parameter sets to guide the weighted fusion of high- and low-resolution features. Accurate and stable quantification of local characteristics enables the weight parameters to accurately reflect the local region's demand for high- and low-resolution features, thereby optimizing the feature fusion process and enhancing the fused feature map's ability to characterize small contaminated targets. This method of accurately quantifying local characteristics, combined with the subsequent weight generation and feature fusion steps, improves the effectiveness of the entire recognition method, especially when identifying small targets under complex imaging conditions.

[0101] More specifically, the above-mentioned calculation method accurately and stably quantifies the local contrast, local texture complexity, and local motion of each image block. These quantization results constitute a multidimensional feature describing the local characteristics of the image block, which is used to generate a set of weight parameters in subsequent steps.

[0102] In some preferred embodiments, the step of dividing the corrected image to be recognized into a plurality of image blocks according to the gradient magnitude and the gradient direction includes:

[0103] S421, setting the weight of the pixel point according to the gradient amplitude, the larger the gradient amplitude, the higher the weight;

[0104] S422: Using a weighted K-means clustering algorithm, with the position coordinates of the pixel points as features and the weights of the pixel points as clustering weights, the corrected image to be identified is divided into multiple image blocks.

[0105] Specifically, the gradient amplitude characterizes the severity of the pixel grayscale change. The gradient amplitude is usually higher at edges, texture-rich areas, or potential target areas. Step S421 sets weights based on the gradient amplitude of the pixel points, so that areas in the image with drastic gradient changes receive higher weights. For example, the weight of the pixel point can be set as a linear function or exponential function of the gradient amplitude. The larger the gradient amplitude, the higher the calculated weight value. This provides an important basis for subsequent image block segmentation and emphasizes the focus on these key areas.

[0106] More specifically, step S422 employs a weighted K-means clustering algorithm to perform image block division. This algorithm uses the spatial coordinates of pixels as clustering features and the gradient magnitude weights set in step S421 as clustering weights. This means that the clustering process considers not only the positional relationship of pixels in the image but also the importance of their gradient magnitudes. This weighted clustering algorithm can cluster pixels with similar gradient magnitudes and spatial proximity into the same image block. Compared to simple grid division or clustering that ignores gradient information, this image block division method better respects the structure and content of the image itself, making the divided image blocks more meaningful and more accurately representing the local contrast, local texture complexity, and local motion within their internal regions. This improved image block division method, combined with the step of calculating local characteristics (such as local contrast, local texture complexity, and local motion), can provide more accurate local information, thereby optimizing the subsequent weight calculation and feature fusion processes and improving overall recognition performance.

[0107] More specifically, through the above technical solution, the present application can divide the corrected image to be identified into multiple image blocks by weighted clustering based on the gradient amplitude and spatial position information of the pixel points. Compared with traditional grid division or clustering methods that do not consider gradient information, this division method can more effectively classify pixels with similar gradient characteristics and adjacent space into the same area, and can better highlight the key areas in the image where the gradient changes dramatically. As a result, local characteristics such as local contrast, local texture complexity and local motion degree calculated based on these image blocks can more accurately reflect the actual content and structure of the image. The calculated local characteristics are more accurate, which provides a more reliable basis for the subsequent weighted fusion of high and low resolution feature maps based on these local characteristics. The weighted fusion process can more accurately adjust the contribution of features of different resolutions according to the importance of local areas of the image, thereby enhancing the ability to integrate image details and robust features.

[0108] In some preferred embodiments, step S41 includes:

[0109] S411, using the Sobel operator to calculate the gradient components of the corrected image to be identified in the horizontal and vertical directions respectively, to obtain a horizontal gradient map and a vertical gradient map;

[0110] S412, calculating the gradient amplitude of each pixel according to the horizontal gradient map and the vertical gradient map, and processing the gradient amplitude using a non-maximum suppression algorithm to obtain a refined gradient amplitude map;

[0111] S413 , calculating the gradient direction of each pixel according to the horizontal gradient map and the vertical gradient map to obtain a discretized gradient direction map.

[0112] Specifically, step S411 uses the Sobel operator to calculate the horizontal and vertical gradient components of the corrected image to be identified, generating horizontal and vertical gradient maps. The Sobel operator is a commonly used edge detection operator. Through convolution operations, it effectively captures areas of image grayscale with dramatic changes, i.e., edge information. It also calculates the horizontal and vertical gradient strength and direction information, respectively, providing basic data for subsequent gradient magnitude and direction calculations.

[0113] More specifically, step S412 calculates the gradient magnitude of each pixel based on the horizontal and vertical gradient maps and processes the gradient magnitude using a non-maximum suppression algorithm to obtain a refined gradient magnitude map. The gradient magnitude represents the severity of the pixel grayscale change, that is, the strength of the edge. The non-maximum suppression algorithm refines the edges and removes false responses by retaining pixels with local maxima along the gradient direction and suppressing pixels with non-maxima. This allows the resulting gradient magnitude map to more accurately reflect the true edge position and strength.

[0114] More specifically, step S413 calculates the gradient direction of each pixel based on the horizontal gradient map and the vertical gradient map, obtaining a discretized gradient direction map. The gradient direction indicates the direction of the pixel grayscale change, i.e., the direction of the edge. Calculating the gradient direction can provide important information about the image texture. Discretizing the gradient direction simplifies the representation while retaining the primary direction information, facilitating the calculation of the gradient direction histogram in the subsequent step S42 as a measure of local texture complexity.

[0115] More specifically, through the above steps, the present application uses the Sobel operator to calculate gradient components, combines the non-maximum suppression algorithm to refine the gradient magnitude, and discretizes the gradient direction, thereby obtaining gradient magnitude and direction information that is more accurate, detailed, and robust than simple gradient calculation. This high-quality gradient information can be more reliably used to calculate local area characteristics in the subsequent step S42, for example, to more accurately measure local contrast and texture complexity.

[0116] In some preferred embodiments, step S2 includes:

[0117] S21, extracting feature points of each image frame in the image sequence and constructing a feature point set;

[0118] S22. For each feature point in the feature point set, a search window is set with the pixel at the corresponding position in the previous frame image as the center. A cross-correlation algorithm is used to perform matching within the search window of the current frame image to obtain the corresponding position of the feature point in the current frame image, and the pixel offset of the feature point between adjacent frames is calculated.

[0119] S23. According to the pixel offset, a global affine transformation matrix is obtained by fitting as a correction parameter for compensating for slight image shaking;

[0120] S24 , using a bilinear interpolation algorithm to perform position correction on the image frames in the image sequence according to the correction parameters, to generate a corrected image sequence including the corrected images to be recognized.

[0121] Specifically, in step S21, feature points are extracted from each frame of the image sequence. These feature points are pixel locations in the image that are unique and stable, such as corner points, edge intersections, or areas with rich textures. The purpose of extracting feature points is to establish a reliable correspondence between different frames as a basis for calculating image motion. A variety of feature point detection algorithms can be used, such as Harris corner detection, Shi-Tomasi corner detection, SIFT, SURF, ORB, etc. The extracted feature points are organized into a feature point set for easy subsequent processing.

[0122] More specifically, in step S22, for each feature point in the feature point set, matching is performed between adjacent frames. A limited search window is set in the current frame with the position of the feature point in the previous frame as the center. The size of the search window can be determined based on the expected maximum pixel offset. Within the search window, a cross-correlation algorithm is used to find the area that is most similar to the image block around the feature point of the previous frame, so as to determine the corresponding position of the feature point in the current frame. The cross-correlation algorithm measures the similarity by calculating the correlation between the two image blocks. The higher the correlation, the higher the similarity. By comparing the position of the feature point in the previous frame and the current frame, its pixel offset between adjacent frames is calculated. These offsets reflect the movement of the image in the local area.

[0123] More specifically, in step S23, a global affine transformation matrix is fitted using the calculated pixel offsets of multiple feature points. Affine transformation is a linear geometric transformation that can describe image movements such as translation, rotation, scaling, and shearing. Through robust fitting methods such as least squares or RANSAC (random sample consensus), the motion information of multiple local feature points can be integrated into a parameter model that describes the overall motion of the entire image, namely the global affine transformation matrix. This global affine transformation matrix is used as a correction parameter specifically for compensating for slight jitters in the image sequence to ensure that subsequent processing is based on a relatively stable image perspective. For example, for slight translations and rotations, affine transformation can provide an accurate description.

[0124] More specifically, in step S24, bilinear interpolation is a commonly used image resampling technique that calculates the grayscale value of the interpolated point based on the grayscale values of four surrounding known pixels and their distances from the interpolated point. Using bilinear interpolation for position correction smoothly transforms the image, avoiding jagged edges and generating a corrected image. These corrected images are combined into a corrected image sequence, which is then used in subsequent steps for feature extraction and object detection. This precise position correction effectively eliminates the effects of minor motion, making subsequent feature extraction more stable and object detection more accurate.

[0125] More specifically, through the above-mentioned technical solution, this application solves the problem of inter-frame misalignment of image sequences caused by minor shakes during image acquisition. Through feature point-based matching and global affine transformation, minute image displacements are accurately calculated and compensated. This position correction process generates a corrected image sequence with inter-frame alignment, providing stable image data for subsequent feature extraction and target detection. This stability supports the accurate acquisition of image features, especially for small targets, reducing feature ambiguity and position uncertainty caused by misalignment, thereby supporting the identification of contaminated targets.

[0126] In some preferred embodiments, step S5 includes:

[0127] S51, for the high-resolution feature map and the low-resolution feature map, respectively constructing a weight parameter matrix corresponding to the corrected image according to the weight parameter group;

[0128] S52, extracting a high-resolution weight parameter of an image block corresponding to each pixel in the high-resolution feature map according to a weight parameter matrix corresponding to the high-resolution feature map, and adjusting a feature vector corresponding to the pixel in the high-resolution feature map according to the high-resolution weight parameter to obtain an adjusted high-resolution feature map;

[0129] S53, extracting a low-resolution weight parameter of an image block corresponding to each pixel in the low-resolution feature map according to a weight parameter matrix corresponding to the low-resolution feature map, and adjusting a feature vector corresponding to the pixel in the low-resolution feature map according to the low-resolution weight parameter to obtain an adjusted low-resolution feature map;

[0130] S54: Concatenate the adjusted high-resolution feature map and the adjusted low-resolution feature map in the channel dimension to obtain a fused feature map.

[0131] Specifically, in step S51 , the weight parameter group includes weight parameters calculated based on a plurality of image blocks divided based on the corrected image, and each image block corresponds to one or a group of weight parameters.

[0132] More specifically, in steps S52 and S53, the weight parameter matrix has the same spatial resolution as the high-resolution feature map and the low-resolution feature map. This can be achieved by assigning the weight parameters of each image block to all pixel positions corresponding to the image block on the feature map, or by upsampling the weight parameters of the image block (for example, using bilinear interpolation) to the resolution of the feature map to construct the weight parameter matrix. Steps S52 and S53 associate the local feature weights of the image domain with each spatial position in the feature domain, solving the problem that the original weight parameter group cannot be directly applied to the feature map.

[0133] More specifically, step S54 combines two different types of feature information, both optimized according to local characteristics, by splicing high- and low-resolution feature maps adjusted by the weight parameter matrix in the channel dimension. This fusion method allows the final feature map to contain both high-frequency details and robust low-frequency information that have been locally enhanced or suppressed. This step, combined with the step of generating a weight parameter group based on local area characteristic analysis, implements weighted adjustment of feature maps based on local image characteristics. This adjustment is performed before feature fusion, allowing the fusion process to utilize the local information of the image to guide the contribution of features of different resolutions, thereby generating a fused feature map that contains richer, more discriminative, and more adaptable to local changes in feature representation, thereby improving the detection accuracy of tiny pollution targets.

[0134] In some preferred embodiments, step S3 includes:

[0135] S31, preprocessing the corrected image to be recognized, where the preprocessing includes image denoising, contrast enhancement, and normalization operations to obtain a preprocessed image;

[0136] S32, inputting the preprocessed image into a preset high-resolution convolutional neural network to extract a shallow high-resolution feature map, where the high-resolution convolutional neural network includes multiple convolutional layers and pooling layers to retain high-frequency detail information of the image;

[0137] S33. Input the preprocessed image into a preset low-resolution convolutional neural network to extract a deep low-resolution feature map. The low-resolution convolutional neural network includes multiple convolutional layers, pooling layers, and downsampling layers for extracting robust low-frequency features of the image.

[0138] Specifically, step S31 pre-processes the corrected image to be identified, aiming to improve image quality and provide better input for subsequent feature extraction.

[0139] More specifically, steps S32 and S33 perform feature extraction on the preprocessed image in parallel.

[0140] More specifically, step S32 feeds the preprocessed image into a pre-set high-resolution convolutional neural network. This network is designed to preserve high-frequency detail in the image and therefore includes multiple convolutional and pooling layers. The convolutional layers extract local features of the image using convolution kernels, while the pooling layers reduce the spatial size of the feature map while retaining key features. The shallow feature maps extracted by this network typically have high spatial resolution and contain abundant high-frequency details such as edges and textures, which are crucial for identifying small contaminated targets that rely on fine structure.

[0141] More specifically, step S33 inputs the preprocessed image into a preset low-resolution convolutional neural network. The design of this network focuses on extracting robust low-frequency features of the image, so in addition to the convolution layer and the pooling layer, it also includes a downsampling layer. The downsampling operation reduces the spatial resolution of the feature map, but can increase the receptive field, enabling the network to capture the global and contextual information of the image. Through multi-layer convolution and downsampling, the deep feature map extracted by the network has a lower spatial resolution, but has better robustness to translation, rotation and scale changes of the image, that is, it extracts robust low-frequency features of the image. This low-frequency feature helps to understand the overall structure and background information of the image and enhances resistance to noise and background changes.

[0142] Therefore, the features output in step S3 combine the feature representation of detail preservation and robustness extraction, overcoming the limitation of a single processing path that is difficult to adapt to complex imaging conditions, and providing more effective information for subsequent feature fusion and target detection, thereby improving the performance of identifying tiny pollution targets in complex and changing environments.

[0143] In some preferred embodiments, step S6 includes:

[0144] S61, performing a convolution operation on the fused feature map and each convolution kernel in a preset convolution kernel set to obtain a plurality of response maps corresponding to pollution types, wherein each convolution kernel in the convolution kernel set performs response enhancement for different types of pollution targets;

[0145] S62. For each response map, a non-maximum suppression algorithm is used for post-processing to screen candidate boxes of contaminated targets with high confidence. The pollution identification result is output based on the position, size, and response intensity of the candidate boxes and the mapping relationship between pollution type and response intensity.

[0146] Specifically, in step S61, the fused feature map combines high-resolution details with robust low-frequency features, containing rich target information. By convolving it with a preset convolution kernel that enhances the response to a specific pollution type, the regions associated with that specific pollution type in the fused feature map can produce a stronger response, thereby highlighting different types of pollution targets and suppressing background interference, laying the foundation for subsequent accurate detection. Each convolution kernel is designed to capture the pattern characteristics of a specific pollution type, allowing the convolution operation to effectively transform the abstract features in the fused feature map into spatial response intensities associated with that specific pollution type.

[0147] More specifically, in step S62, non-maximum suppression is applied to the response map to extract candidate boxes representing potential contamination targets. This ensures that each target is detected only once, improving the cleanliness and accuracy of the detection results. The final contamination identification result is output based on the position, size, and response strength of the selected candidate boxes, combined with a preset mapping between contamination type and response strength.

[0148] More specifically, through the above technical solution, this application solves the problem of the difficulty in accurately identifying small polluted targets under complex dynamic imaging conditions when performing general target detection directly based on fused feature maps. By convolving the fused feature map with a convolution kernel preset for a specific pollution type, the response of small polluted targets in the feature map can be effectively enhanced, while suppressing the interference of background noise and artifacts, allowing different types of polluted targets to stand out.

[0149] Second, please refer to Figure 2 Some embodiments of the present application further provide a convolutional neural network-based polypeptide drug membrane filtration contamination identification system, the system comprising:

[0150] An acquisition module 201 is used to acquire an image to be identified on the surface of the filter membrane and a corresponding image sequence;

[0151] A correction module 202 is configured to perform position correction on the image sequence to generate a corrected image sequence including corrected images to be recognized;

[0152] The processing module 203 is configured to input the corrected image to be recognized in the corrected image sequence into a preset high-resolution processing path and a preset low-resolution processing path to obtain a high-resolution feature map and a low-resolution feature map;

[0153] A weight module 204 is used to analyze the local area characteristics of the corrected image to be identified to generate a weight parameter group;

[0154] A fusion module 205 is configured to perform weighted fusion of the high-resolution feature map and the low-resolution feature map according to the weight parameter group to generate a fused feature map;

[0155] The identification module 206 is configured to perform target detection based on the fused feature map to generate a pollution identification result.

[0156] The core innovation of the polypeptide drug membrane filtration pollution identification system of the convolutional neural network in the embodiment of the present application lies in the combination of high- and low-resolution dual-path processing and weighted fusion based on local area characteristics, thereby generating a fusion feature map for target detection, effectively dealing with the problems of slight shaking and detail loss under complex imaging conditions, thereby achieving accurate identification of tiny pollution targets, having better robustness, being able to effectively identify tiny pollution targets, and reducing false alarm and missed alarm rates.

[0157] In addition, the units described as separate components may or may not be physically separate, and 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 units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0158] Furthermore, the functional modules in each embodiment of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0159] In this document, relational terms such as first and second, etc. are used merely to distinguish one entity or operation from another entity or operation, but do not necessarily require or imply any actual relationship or order between these entities or operations.

[0160] The above description is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. For those skilled in the art, various modifications and variations of the present application are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A method for identifying polypeptide drug membrane filtration contamination using a convolutional neural network, characterized in that: The method comprises the following steps: S1. Obtain the image to be identified and the corresponding image sequence on the surface of the filter membrane; S2. performing position correction on the image sequence to generate a corrected image sequence including corrected images to be recognized; S3, inputting the corrected image to be recognized in the corrected image sequence into a preset high-resolution processing path and a preset low-resolution processing path to obtain a high-resolution feature map and a low-resolution feature map; S4, analyzing the local area characteristics of the corrected image to be identified to generate a weight parameter group; S5. Weightedly fuse the high-resolution feature map and the low-resolution feature map according to the weight parameter group to generate a fused feature map; S6. Performing target detection based on the fused feature map to generate a pollution identification result; The local area characteristics include local contrast, local texture complexity and local motion degree. Step S4 includes: S41, calculating the gradient magnitude and gradient direction of each pixel in the corrected image to be identified; S42, dividing the corrected image to be recognized into a plurality of image blocks according to the gradient magnitude and the gradient direction, and calculating the local contrast, local texture complexity, and local motion degree of each image block; S43, calculating initial weight parameters corresponding to the high-resolution processing path and the low-resolution processing path for each image block using a preset nonlinear mapping model based on the local contrast, local texture complexity, and local motion degree of each image block; S44: Normalize the initial weight parameters to obtain a final weight parameter group.

2. The method for identifying polypeptide drug membrane filtration pollution using a convolutional neural network according to claim 1, characterized in that: Step S43 includes: S431, for each image block, constructing a multidimensional feature vector including local contrast, local texture complexity, and local motion degree; S432. Input the multidimensional feature vector into a pre-trained nonlinear mapping model to output initial weight parameters corresponding to a high-resolution processing path and a low-resolution processing path for each image block, wherein the initial weight parameters of the high-resolution processing path are positively correlated with the local contrast and the local texture complexity, and negatively correlated with the local motion degree.

3. The method for identifying polypeptide drug membrane filtration pollution using a convolutional neural network according to claim 1, characterized in that: The steps of calculating the local contrast, local texture complexity and local motion degree of each image block include: S423, calculating the standard deviation of the grayscale values of all pixels in each image block, and dividing the standard deviation by the mean grayscale values of all pixels in the image block to obtain a normalized local contrast; S424, calculating the histogram of the gradient direction of all pixels in each image block, counting the number of pixels in each gradient direction, and calculating the entropy value of the histogram as the local texture complexity; S425 , calculating and obtaining the optical flow vector between each image block and the corresponding image block of the previous frame, and calculating the average amplitude of the optical flow vector to obtain the degree of local motion.

4. The method for identifying polypeptide drug membrane filtration pollution using a convolutional neural network according to claim 1, characterized in that: The step of dividing the corrected image to be recognized into a plurality of image blocks according to the gradient magnitude and the gradient direction comprises: S421, setting the weight of the pixel point according to the gradient amplitude, the larger the gradient amplitude, the higher the weight; S422: Using a weighted K-means clustering algorithm, with the position coordinates of the pixel points as features and the weights of the pixel points as clustering weights, the corrected image to be identified is divided into multiple image blocks.

5. The method for identifying polypeptide drug membrane filtration pollution using a convolutional neural network according to claim 1, characterized in that: Step S41 includes: S411, using the Sobel operator to calculate the gradient components of the corrected image to be identified in the horizontal and vertical directions respectively, to obtain a horizontal gradient map and a vertical gradient map; S412, calculating the gradient amplitude of each pixel according to the horizontal gradient map and the vertical gradient map, and processing the gradient amplitude using a non-maximum suppression algorithm to obtain a refined gradient amplitude map; S413 , calculating the gradient direction of each pixel according to the horizontal gradient map and the vertical gradient map to obtain a discretized gradient direction map.

6. The method for identifying polypeptide drug membrane filtration pollution using a convolutional neural network according to claim 1, characterized in that: Step S5 includes: S51, for the high-resolution feature map and the low-resolution feature map, respectively constructing a weight parameter matrix corresponding to the corrected image according to the weight parameter group; S52, extracting a high-resolution weight parameter of an image block corresponding to each pixel in the high-resolution feature map according to a weight parameter matrix corresponding to the high-resolution feature map, and adjusting a feature vector corresponding to the pixel in the high-resolution feature map according to the high-resolution weight parameter to obtain an adjusted high-resolution feature map; S53, extracting a low-resolution weight parameter of an image block corresponding to each pixel in the low-resolution feature map according to a weight parameter matrix corresponding to the low-resolution feature map, and adjusting a feature vector corresponding to the pixel in the low-resolution feature map according to the low-resolution weight parameter to obtain an adjusted low-resolution feature map; S54: Concatenate the adjusted high-resolution feature map and the adjusted low-resolution feature map in the channel dimension to obtain a fused feature map.

7. The method for identifying polypeptide drug membrane filtration pollution using a convolutional neural network according to claim 1, characterized in that: Step S3 includes: S31, preprocessing the corrected image to be recognized to obtain a preprocessed image; S32, inputting the preprocessed image into a preset high-resolution convolutional neural network to extract a high-resolution feature map, wherein the high-resolution convolutional neural network includes multiple convolutional layers and pooling layers; S33. Input the preprocessed image into a preset low-resolution convolutional neural network to extract a low-resolution feature map. The low-resolution convolutional neural network includes multiple convolution layers, pooling layers and downsampling layers.

8. The method for identifying polypeptide drug membrane filtration pollution using a convolutional neural network according to claim 1, characterized in that: Step S6 includes: S61, performing a convolution operation on the fused feature map and each convolution kernel in a preset convolution kernel set to obtain a plurality of response maps corresponding to pollution types, wherein each convolution kernel in the convolution kernel set performs response enhancement for different types of pollution targets; S62. For each response map, a non-maximum suppression algorithm is used for post-processing to screen candidate boxes of contaminated targets with high confidence. The pollution identification result is output based on the position, size, and response intensity of the candidate boxes and the mapping relationship between pollution type and response intensity.

9. A convolutional neural network polypeptide drug membrane filtration pollution identification system, characterized in that: The system comprises: An acquisition module, used to acquire the image to be identified and the corresponding image sequence on the surface of the filter membrane; a correction module, configured to perform position correction on the image sequence to generate a corrected image sequence containing corrected images to be recognized; a processing module, configured to input the corrected image to be identified in the corrected image sequence into a preset high-resolution processing path and a preset low-resolution processing path to obtain a high-resolution feature map and a low-resolution feature map; A weight module, used for analyzing the local area characteristics of the corrected image to be identified to generate a weight parameter group; A fusion module, configured to perform weighted fusion of the high-resolution feature map and the low-resolution feature map according to the weight parameter group to generate a fused feature map; an identification module, configured to perform target detection based on the fused feature map to generate a pollution identification result; The local area characteristics include local contrast, local texture complexity, and local motion degree. The process of analyzing the corrected local area characteristics of the image to be identified to generate a weight parameter group includes: S41, calculating the gradient magnitude and gradient direction of each pixel in the corrected image to be identified; S42, dividing the corrected image to be recognized into a plurality of image blocks according to the gradient magnitude and the gradient direction, and calculating the local contrast, local texture complexity, and local motion degree of each image block; S43, calculating initial weight parameters corresponding to the high-resolution processing path and the low-resolution processing path for each image block using a preset nonlinear mapping model based on the local contrast, local texture complexity, and local motion degree of each image block; S44: Normalize the initial weight parameters to obtain a final weight parameter group.

Images