Polypeptide drug membrane filtration pollution identification method and system based on convolutional neural network

Abstract

The invention relates to the technical field of target recognition, and particularly discloses a polypeptide drug membrane filtration pollution recognition method and system based on a convolutional neural network, and the method comprises the steps: obtaining a to-be-recognized image and a corresponding image sequence; generating a corrected image sequence containing the corrected to-be-recognized image; inputting the corrected to-be-recognized image in the corrected image sequence to obtain a high-resolution feature map and a low-resolution feature map; analyzing local region characteristics of the corrected to-be-recognized image to generate a weight parameter group; weighting and 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; according to the method, the problems of tiny shaking and detail loss under the complex imaging condition are effectively solved, so that accurate recognition of the tiny pollution target is achieved, better robustness is achieved, the tiny pollution target can be effectively recognized, and the false alarm rate and the missing report rate are reduced.

Description

Technical Field

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

[0002] In the production of peptide drugs, membrane filtration is a key purification step to remove impurities to ensure product quality and drug safety. During production, the membrane filtration equipment operates continuously, and the purified solution is pumped through a filter 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. An image acquisition unit is installed at a key position of the equipment, including a high-resolution camera and a light source, to periodically or continuously acquire image data and transmit it to a processing unit, which is analyzed by an image recognition program based on a convolutional neural network (CNN) to determine contamination, identify type, determine location, estimate size, and record information, triggering an alarm or pausing the process when necessary. This method improves detection efficiency, consistency, and objectivity. However, actual imaging conditions are complex and dynamically changing, and CNN models based on idealized training data are difficult to adapt to. It is difficult to identify tiny contaminated targets, and false alarms or missed alarms are prone to occur, which 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, so as to cope with the problems of slight shaking and detail loss under complex imaging conditions, accurately identify tiny contaminated 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: S1, obtaining the image to be identified on the surface of the filter membrane and the corresponding image sequence; 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 fusing the high-resolution feature map and the low-resolution feature map according to the weight parameter group to generate a fused feature map; S6. Perform object detection based on the fused feature map to generate a pollution recognition result.

[0007] The core innovation of the polypeptide drug membrane filtration pollution recognition method using the convolutional neural network in this application lies in combining high - and low - resolution dual - path processing and weighted fusion based on local region characteristics, thereby generating a fused feature map for object detection, effectively addressing the problems of slight shaking and detail loss under complex imaging conditions, thus achieving accurate recognition of tiny pollution targets, having better robustness, being able to effectively identify tiny pollution targets, and reducing the false alarm and miss rate.

[0008] In the described polypeptide drug membrane filtration pollution recognition method using a convolutional neural network, the local region characteristics include local contrast, local texture complexity, and local motion degree. Step S4 includes: S41. Calculate the gradient magnitude and gradient direction of each pixel point in the corrected image to be recognized. S42. According to the gradient magnitude and the gradient direction, divide the corrected image to be recognized into multiple image blocks, and calculate the local contrast, local texture complexity, and local motion degree of each image block. S43. According to the local contrast, local texture complexity, and local motion degree of each image block, use a preset non - linear mapping model to calculate the initial weight parameters corresponding to the high - resolution processing path and the low - resolution processing path for each image block. S44. Perform normalization processing on the initial weight parameters to obtain the final weight parameter group.

[0009] Through the above technical solution, this application can generate a more refined and more discriminative weight parameter group according to the specific characteristics such as the local contrast, local texture complexity, and local motion degree of the corrected image to be recognized.

[0010] In the described polypeptide drug membrane filtration pollution recognition method using a convolutional neural network, step S43 includes: S431. For each image block, construct a multi - dimensional feature vector including local contrast, local texture complexity, and local motion degree. S432. Input the multi - dimensional feature vector into a pre - trained non - linear mapping model to output the 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 local contrast and local texture complexity and negatively correlated with local motion degree.

[0011] In the described polypeptide drug membrane filtration pollution recognition method using a convolutional neural network, the steps of calculating the local contrast, local texture complexity, and local motion degree of each image block include: S423. Calculate the standard deviation of the gray values of all pixels within each image patch, and divide it by the mean of the gray values of all pixels within the image patch to obtain the normalized local contrast; S424. Calculate the histogram of the gradient directions of all pixel points within each image patch, count the number of pixel points in each gradient direction, and calculate the entropy value of the histogram as the local texture complexity; S425. Calculate and obtain the optical flow vector between each image patch and the corresponding image patch in the previous frame, and count the average magnitude of the optical flow vectors to obtain the local motion degree.

[0012] The described method for identifying polypeptide drug membrane filtration contamination by a convolutional neural network, wherein the step of dividing the corrected image to be identified into multiple image patches according to the gradient magnitude and the gradient direction includes: S421. Set the weights of pixel points according to the gradient magnitude, and the greater the gradient magnitude, the higher the weight; S422. Use the weighted K-means clustering algorithm, with the position coordinates of pixel points as features and the weights of pixel points as clustering weights, to divide the corrected image to be identified into multiple image patches.

[0013] The described method for identifying polypeptide drug membrane filtration contamination by a convolutional neural network, wherein step S41 includes: S411. Use the Sobel operator to calculate the gradient components of the corrected image to be identified in the horizontal and vertical directions respectively to obtain the horizontal gradient map and the vertical gradient map; S412. According to the horizontal gradient map and the vertical gradient map, calculate the gradient magnitude of each pixel point, and use the non-maximum suppression algorithm to process the gradient magnitude to obtain a refined gradient magnitude map; S413. According to the horizontal gradient map and the vertical gradient map, calculate the gradient direction of each pixel point to obtain a discretized gradient direction map.

[0014] The described method for identifying polypeptide drug membrane filtration contamination by a convolutional neural network, wherein step S5 includes: S51. For the high-resolution feature map and the low-resolution feature map, respectively construct weight parameter matrices corresponding to the corrected image according to the weight parameter group; S52. According to the weight parameter matrix corresponding to the high-resolution feature map, extract the high-resolution weight parameters of the image patches corresponding to each pixel point in the high-resolution feature map, and adjust the feature vectors of the corresponding pixel points in the high-resolution feature map according to the high-resolution weight parameters to obtain an adjusted high-resolution feature map; S53. Extract the low-resolution weight parameters of the image patches corresponding to each pixel in the low-resolution feature map according to the weight parameter matrix corresponding to the low-resolution feature map, and adjust the feature vectors of the corresponding pixels in the low-resolution feature map according to the low-resolution weight parameters 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.

[0015] In the described method for identifying polypeptide drug membrane filtration contamination by a convolutional neural network, step S3 includes: S31. Preprocess the corrected image to be recognized to obtain a preprocessed image. S32. Input the preprocessed image into a preset high-resolution convolutional neural network to extract a high-resolution feature map. 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 convolutional layers, pooling layers, and downsampling layers.

[0016] In the described method for identifying polypeptide drug membrane filtration contamination by a convolutional neural network, step S6 includes: S61. Perform a convolution operation on the fused feature map with each convolution kernel in a preset set of convolution kernels to obtain multiple response maps corresponding to the contamination types. Each convolution kernel in the set of convolution kernels enhances the response to different types of contamination targets. S62. For each response map, perform post-processing using the non-maximum suppression algorithm to screen the candidate boxes of contamination targets with high confidence. According to the positions, sizes, and response intensities of the candidate boxes, and in combination with the mapping relationship between the contamination type and the response intensity, output the contamination recognition result.

[0017] In a second aspect, the present application also provides a system for identifying polypeptide drug membrane filtration contamination by a convolutional neural network. The system includes: An acquisition module for acquiring the image to be recognized on the surface of the filter membrane and the corresponding image sequence. A correction module for performing position correction on the image sequence to generate a corrected image sequence including the corrected image to be recognized. A processing module for inputting the corrected image to be recognized in the corrected image sequence into a preset high-resolution processing path and a low-resolution processing path to obtain a high-resolution feature map and a low-resolution feature map. A weight module for analyzing the local region characteristics of the corrected image to be recognized to generate a set of weight parameters. A fusion module, configured to generate a fused feature map by fusing the high-resolution feature map and the low-resolution feature map according to the weight parameter group; An identification module, configured to perform target detection based on the fused feature map to generate a pollution identification result.

[0018] The polypeptide drug membrane filtration pollution identification system of the convolutional neural network of the present application combines high and low resolution dual-path processing and weighted fusion based on local region characteristics, thereby generating a fused feature map for target detection, effectively coping with the problems of slight shaking and detail loss under complex imaging conditions, and thus realizing accurate identification of tiny pollution targets.

[0019] As can be seen from the above, the present application provides a method and system for identifying polypeptide drug membrane filtration pollution of a convolutional neural network. Among them, the method of the present application solves the problems of poor robustness, easy false alarms and missed detections of existing convolutional neural network methods when identifying tiny pollution targets of polypeptide drug membrane filtration under complex dynamic imaging conditions. By performing position correction on the image sequence, it eliminates the interference of image shaking on identification and improves the stability of processing. By parallel processing of high-resolution and low-resolution paths, it captures the detailed information important for tiny targets and the low-frequency features robust to environmental changes respectively. By analyzing local region characteristics and performing weighted fusion, the feature representation can be optimized according to the actual situation of different regions of the image, enhancing the sensitivity to tiny targets and at the same time improving the adaptability to complex imaging conditions. Based on the optimized fused feature map for target detection, it can more accurately distinguish real pollution targets from background interference, thereby improving the identification accuracy and robustness of tiny pollution targets and reducing the false alarm and missed detection rates. Description of the Drawings

[0020] Figure 1 It is a flowchart of a method for identifying polypeptide drug membrane filtration pollution of a convolutional neural network provided by an embodiment of the present application.

[0021] Figure 2 It is a schematic structural diagram of a polypeptide drug membrane filtration pollution identification system of a convolutional neural network provided by an embodiment of the present application.

[0022] Reference numerals: 201, an acquisition module; 202, a correction module; 203, a processing module; 204, a weight module; 205, a fusion module; 206, an identification module. Detailed Embodiments

[0023] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying 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 the embodiments. Usually, the components of the embodiments of the present application described and shown in the accompanying 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 accompanying drawings is not intended to limit the scope of the present application claimed, but only represents the selected embodiments of the present application. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative efforts belong to the scope of protection of the present application.

[0024] It should be noted that similar reference numerals and letters denote similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of the present application, terms such as "first" and "second" are only used for differential description and cannot be understood as indicating or implying relative importance.

[0025] In a first aspect, please refer to Figure 1 , some embodiments of the present application provide a method for identifying polypeptide drug membrane filtration pollution in a convolutional neural network. The method includes the following steps: S1. Obtain the image to be recognized on the surface of the filter membrane and the corresponding image sequence; S2. Perform position correction on the image sequence to generate a corrected image sequence including the corrected image to be recognized; S3. Input the corrected image to be recognized in the corrected image sequence into a preset high-resolution processing path and a low-resolution processing path to obtain a high-resolution feature map and a low-resolution feature map; S4. Analyze the local region characteristics of the corrected image to be recognized 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. Perform object detection based on the fused feature map to generate a pollution recognition result.

[0026] Specifically, position correction refers to adjusting the spatial positions of the images in each frame of the image sequence to align them with each other. Specifically, it can be achieved by analyzing the pixel offsets between the image frames and applying transformations. Through position correction, the slight shaking or drift during the image acquisition process is compensated, ensuring the spatial alignment between the frames in the image sequence, ensuring the spatial consistency of the images in the sequence, and providing a stable input for subsequent feature extraction and analysis.

[0027] More specifically, the high-resolution processing path and the low-resolution processing path refer to inputting an image into two different computational processes. The high-resolution processing path preserves the high-frequency details of the image, and the low-resolution processing path extracts the robust low-frequency features of the image. Therefore, the high-resolution processing path extracts and preserves the features with fine details of the image, and the low-resolution processing path extracts the features with adaptability to image changes, that is, the feature representations sensitive to details and adaptable to changes are obtained simultaneously. The two paths can be specifically implemented using different convolutional neural network structures.

[0028] More specifically, the local region feature analysis refers to calculating the attributes of each small region in the image. These attributes can include contrast, texture, and motion information. Specifically, it can be implemented by calculating the statistics within the region. This provides information about the content and nature of different regions of the image.

[0029] More specifically, the weight parameter group refers to a set of numerical values generated according to the results of the local region feature analysis. These numerical values are used to guide the subsequent feature fusion process.

[0030] More specifically, the weighted fusion refers to combining the high-resolution feature map and the low-resolution feature map according to the weight parameter group. Specifically, it can be implemented by applying the weights to the feature maps and then performing splicing or summation. This adaptively integrates different types of features according to the characteristics of the local regions of the image, generating a representation containing more comprehensive information.

[0031] More specifically, the target detection refers to identifying, locating, and determining the size and type of the pollution target on the fused feature map. Specifically, it can be implemented by convolving the convolutional kernel with the feature map and then performing post-processing. This outputs the final pollution recognition result, including the position, size, and type of the tiny pollution target.

[0032] More specifically, the method of this application solves the problems of poor robustness, easy false alarms and missed detections of existing convolutional neural network methods when identifying tiny pollution targets in polypeptide drug membrane filtration under complex dynamic imaging conditions. It eliminates the interference of image shaking on recognition by performing position correction on the image sequence, improving the stability of processing. By processing the high-resolution and low-resolution paths in parallel, it captures the detail information important for tiny targets and the low-frequency features robust to environmental changes respectively. By analyzing the local region characteristics and performing weighted fusion, the feature representation can be optimized according to the actual situation of different regions of the image, enhancing the sensitivity to tiny targets and at the same time improving the adaptability to complex imaging conditions. Based on the optimized fused feature map for target detection, it can more accurately distinguish real pollution targets from background interference, thereby improving the recognition accuracy and robustness of tiny pollution targets and reducing the false alarm and missed detection rates.

[0033] Therefore, the core innovation of the method for identifying polypeptide drug membrane filtration pollution in the convolutional neural network according to the embodiments of the present application lies in combining high- and low-resolution dual-path processing and weighted fusion based on local region characteristics, so as to generate a fused feature map for target detection, effectively cope 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 the false alarm and missed alarm rates.

[0034] In some preferred embodiments, the local region characteristics include local contrast, local texture complexity, and local motion degree, and step S4 includes: S41. Calculate the gradient magnitude and gradient direction of each pixel point in the corrected image to be identified. S42. According to the gradient magnitude and gradient direction, divide the corrected image to be identified into multiple image blocks, and calculate the local contrast, local texture complexity, and local motion degree of each image block. S43. According to the local contrast, local texture complexity, and local motion degree of each image block, use a preset non-linear mapping model to calculate the initial weight parameters corresponding to the high-resolution processing path and the low-resolution processing path for each image block. S44. Perform normalization processing on the initial weight parameters to obtain the final weight parameter group.

[0035] Specifically, in step S41, the gradient magnitude characterizes the severity of pixel gray-scale change, and is related to the edges, details, and contrast of the image; the gradient direction indicates the direction of gray-scale change and is related to the texture structure of the image.

[0036] More specifically, in step S42, dividing the image blocks according to the gradient information helps to group regions with similar structures or change characteristics into the same block. Within each image block, calculate the local contrast, local texture complexity, and local motion degree. These specific local characteristics quantify the properties of the image block from different dimensions, laying a foundation for generating targeted weight parameters subsequently.

[0037] More specifically, in step S43, the non-linear mapping model can be composed of a deep neural network to fit the non-linear relationship between the multi-dimensional feature vector and the weights of the high- and low-resolution processing paths. Calculating the weight parameters for the high-resolution and low-resolution paths respectively is because different local characteristics have different requirements for features of different resolutions: regions with high contrast or complex textures may require higher weights to be assigned to the high-resolution features that retain details, while motion regions or flat regions may rely more on robust low-resolution features. Calculating the weights based on the specific local characteristics enables the weight parameters to more accurately reflect the demand preferences of each local region for high- and low-resolution features.

[0038] More specifically, in step S44, the normalization process ensures that the weight parameters are within a reasonable range, facilitating subsequent weighted fusion operations and ensuring the stability of the fusion process.

[0039] More specifically, through the above technical solution, the present application can generate a more refined and discriminative set of weight parameters according to specific characteristics such as the local contrast, local texture complexity, and local motion degree of the corrected image to be recognized. Thus, in the subsequent feature fusion step, the contribution ratios of the high-resolution feature map and the low-resolution feature map can be adjusted more accurately according to these weight parameters, enabling the fused feature map to better retain the detail information of small targets while suppressing background interference and motion artifacts. This helps to improve the accuracy of subsequent object detection, enabling small contaminated targets to be effectively distinguished and recognized, and reducing false alarms and missed detections.

[0040] In some preferred embodiments, step S43 includes: S431. For each image block, construct a multi-dimensional feature vector including local contrast, local texture complexity, and local motion degree; S432. Input the multi-dimensional feature vector into a pre-trained non-linear mapping model to output the 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 local motion degree.

[0041] Specifically, the above processing method aims to provide a more accurate and robust way to calculate the weight parameters for high-low resolution feature fusion to overcome the deficiencies of simply preset non-linear functions in dealing with the relationship between complex local region characteristics and weights.

[0042] More specifically, step S431 quantifies and integrates the local region characteristics of the image block into a unified input form. The local contrast, local texture complexity, and local motion degree respectively 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 rich and structured input for subsequent weight calculation.

[0043] More specifically, in step S432, the local region feature vector is input into a pre-trained model, which can directly output the weights that the image patch should be assigned on the high-resolution path and the low-resolution path according to the complex mapping relationship it has learned. Among them, the non-linear mapping model is composed of a deep neural network, and the non-linear mapping model fits the non-linear relationship between the multi-dimensional 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 the local texture complexity, and negatively correlated with the local motion degree. This is because the high-resolution path focuses on retaining details, and regions with high contrast and complex textures usually contain rich detail information, while motion may cause detail blurring; the weight of the low-resolution path is negatively correlated with the local contrast and the local texture complexity, and positively correlated with the local motion degree. This is because the low-resolution path focuses on extracting robust features, which are not sensitive to details but may be more stable in regions with motion or where details are not obvious. By clarifying this correlation, it provides a guiding direction for the training of the model.

[0044] More specifically, the non-linear mapping model designed by the deep neural network has a powerful non-linear fitting ability, which can learn and represent the complex relationship between the local region features and the weights that is difficult to describe with a simple function, thus overcoming the limitations of simple pre-set non-linear functions. Compared with calculating weights using simple pre-set functions, using a pre-trained deep neural network can learn and adapt to more complex relationships between local features and weights, thereby generating more accurate weight parameters, significantly improving the effect of feature fusion and the final pollution recognition performance. Especially when dealing with complex and variable actual imaging conditions and identifying small targets, its advantages are more obvious.

[0045] More specifically, through non-linear mapping using a deep neural network, the calculated initial weight parameters can more accurately reflect the requirements of the local features of the current image patch for high and low-resolution feature fusion. After these more accurate initial weight parameters are normalized, they are used to weighted fuse the high-resolution feature map and the low-resolution feature map. This improved weight calculation method makes feature fusion more effective, can better retain important detail information, and at the same time suppress the influence of noise and motion blur. Especially for the recognition of small pollution targets, the detail information is crucial, and accurate weight assignment can significantly improve the recognition accuracy and robustness.

[0046] In some preferred embodiments, the steps of calculating the local contrast, the local texture complexity, and the local motion degree of each image patch include: S423. Calculate the standard deviation of all pixel gray values within each image patch and divide it by the mean of all pixel gray values within the image patch to obtain the normalized local contrast; S424. Calculate the histogram of the gradient directions of all pixel points within each image patch, count the number of pixel points in each gradient direction, and calculate the entropy value of the histogram as the local texture complexity. S425. Calculate and obtain the optical flow vector between each image patch and the corresponding image patch in the previous frame, and count the average magnitude of the optical flow vectors to obtain the local motion degree.

[0047] Specifically, during the calculation of local contrast, the standard deviation measures the degree of dispersion of pixel gray values, reflects the local contrast, and is normalized by dividing by the mean value to eliminate the influence of the overall brightness of the image patch on the contrast measurement and make the result stable.

[0048] More specifically, during the calculation of local texture complexity, the gradient direction histogram describes the direction distribution of edges and textures within the image patch. The entropy value measures the uncertainty of information and represents the degree of uniformity of the gradient direction distribution here. The higher the entropy value, the more uniform the gradient direction distribution and the more complex the texture.

[0049] More specifically, during the calculation of local motion degree, the optical flow vector reflects the motion information of pixels between consecutive frames, and its magnitude represents the motion speed. Calculating the average magnitude can quantify the overall motion intensity of the local area. For example, optical flow algorithms such as Lucas-Kanade can be used to calculate the optical flow vectors of pixel points within the image patch, and then calculate the average value of the magnitudes of these vectors.

[0050] More specifically, the quantization results of local contrast, local texture complexity, and local motion degree obtained through these specific methods are accurate and stable. These quantization results are used to generate a weight parameter group in the subsequent steps to guide the weighted fusion of high- and low-resolution features. Accurate and stable quantization of local characteristics can enable the weight parameters to accurately reflect the needs of the local area for high- and low-resolution features, thereby optimizing the feature fusion process and enhancing the representation ability of the fused feature map for tiny 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 recognizing tiny targets under complex imaging conditions.

[0051] More specifically, through the above specific calculation methods, the local contrast, local texture complexity, and local motion degree of each image patch are accurately and stably quantified. These quantization results constitute multi-dimensional features describing the local characteristics of the image patch and are used to generate a weight parameter group in the subsequent steps.

[0052] In some preferred embodiments, the step of dividing the corrected image to be recognized into multiple image patches according to the gradient magnitude and gradient direction includes: S421. Set the weights of pixel points according to the gradient magnitude, and the greater the gradient magnitude, the higher the weight. S422, using a weighted K-means clustering algorithm, taking 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 a plurality of image blocks.

[0053] Specifically, the gradient amplitude characterizes the drastic degree of pixel grayscale change, and the gradient amplitude is usually higher at the edge, in a texture-rich area, or in a potential target area. Step S421 sets weights according to the gradient amplitude of the pixel points so that the areas with drastic gradient changes in the image obtain higher weights. For example, the weights of the pixels can be set as a linear function or an exponential function of the gradient amplitude. The larger the gradient amplitude, the higher the calculated weight value. This provides an important basis for the subsequent image block division and emphasizes the attention paid to these key areas.

[0054] More specifically, step S422 uses a weighted K-means clustering algorithm to divide the image blocks. The algorithm uses the spatial position coordinates of the pixel points as clustering features, and uses the gradient amplitude weight set in step S421 as the clustering weight. This means that the clustering process not only considers the positional relationship of the pixel points in the image, but also considers the importance of its gradient amplitude. The weighted clustering algorithm can cluster pixels with similar gradient amplitudes and spatially adjacent to the same image block. Compared with simple grid division or clustering without considering gradient information, this image block division method can better respect the structure and content of the image itself, making the divided image blocks more practical and more accurately representing the local contrast, local texture complexity and local motion degree of its internal area. This improved image block division method combined with the step of calculating local characteristics (such as local contrast, local texture complexity and local motion degree) can provide more accurate local information, thereby optimizing the subsequent weight calculation and feature fusion process, and improving the overall recognition performance.

[0055] 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 according to the gradient amplitude and spatial position information of the pixel points. Compared with the traditional grid division or clustering method that does not consider the gradient information, this division method can more effectively classify the pixels with similar gradient characteristics and spatially adjacent into the same area, and can better highlight the key areas in the image where the gradient changes dramatically. Therefore, the 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 according to these local characteristics. The weighted fusion process can more accurately adjust the contribution of different resolution features according to the importance of the local area of ​​the image, thereby enhancing the ability to integrate image details and robust features.

[0056] In some preferred embodiments, step S41 includes: S411. Calculate the gradient components of the corrected image to be recognized in the horizontal and vertical directions respectively using the Sobel operator, to obtain a horizontal gradient map and a vertical gradient map; S412. Calculate the gradient magnitude of each pixel point according to the horizontal gradient map and the vertical gradient map, and process the gradient magnitude using the non-maximum suppression algorithm to obtain a refined gradient magnitude map; S413. Calculate the gradient direction of each pixel point according to the horizontal gradient map and the vertical gradient map, to obtain a discretized gradient direction map.

[0057] Specifically, in step S411, the Sobel operator is used to calculate the gradient components of the corrected image to be recognized in the horizontal and vertical directions, to obtain a horizontal gradient map and a vertical gradient map. The Sobel operator is a commonly used edge detection operator. Through convolution operations, it can effectively capture the regions where the pixel gray values change drastically in the image, that is, edge information, and calculate the gradient intensity and direction information in the horizontal and vertical directions respectively, providing basic data for subsequent gradient magnitude and direction calculations.

[0058] More specifically, in step S412, the gradient magnitude of each pixel point is calculated according to the horizontal gradient map and the vertical gradient map, and the non-maximum suppression algorithm is used to process the gradient magnitude to obtain a refined gradient magnitude map. The gradient magnitude represents the degree of drastic change in pixel gray values, that is, the intensity of the edge. The non-maximum suppression algorithm refines the edge and removes false responses by retaining the pixel points with local maxima along the gradient direction and suppressing the non-maximum pixel points, so that the obtained gradient magnitude map can more accurately reflect the true edge position and intensity.

[0059] More specifically, in step S413, the gradient direction of each pixel point is calculated according to the horizontal gradient map and the vertical gradient map, to obtain a discretized gradient direction map. The gradient direction indicates the direction of pixel gray value change, that is, the direction of the edge. Calculating the gradient direction can provide important information about the image texture. Discretizing the gradient direction can simplify the representation while retaining the main direction information, facilitating the calculation of the gradient direction histogram as a measure of local texture complexity in subsequent step S42.

[0060] More specifically, through the above steps, the present application calculates the gradient components using the Sobel operator, combines the non-maximum suppression algorithm to refine the gradient magnitude, and discretely represents the gradient direction, obtaining more accurate, refined and robust gradient magnitude and direction information than simple gradient calculations. These high-quality gradient information can be more reliably used for the calculation of local region characteristics in subsequent step S42, such as more accurately measuring local contrast and texture complexity.

[0061] In some preferred embodiments, step S2 includes: S21. Extract feature points of each image frame in the image sequence and construct a feature point set; S22. For each feature point in the feature point set, set a search window centered on the pixel point at the corresponding position in the previous frame image. Within the search window of the current frame image, use the cross-correlation algorithm for matching to obtain the corresponding position of the feature point in the current frame image, and calculate the pixel offset of the feature point between adjacent frames; S23. According to the pixel offset, fit to obtain a global affine transformation matrix as the correction parameter for compensating for the slight shaking of the image; S24. According to the correction parameter, use the bilinear interpolation algorithm to perform position correction on the image frames in the image sequence, and generate a corrected image sequence including the corrected image to be recognized.

[0062] Specifically, in step S21, feature points are extracted from each frame of the image sequence. These feature points are pixel positions in the image with uniqueness and stability, such as corner points, edge intersection points, or regions with rich texture. Extracting feature points is to establish a reliable correspondence between different frames as the basis for calculating the movement of the image. Various feature point detection algorithms can be used, such as Harris corner detection, Shi-Tomasi corner detection, SIFT, SURF, ORB and other algorithms. The extracted feature points are organized into a feature point set for subsequent processing.

[0063] More specifically, in step S22, for each feature point in the feature point set, matching is performed between adjacent frames. Centered on the position of the feature point in the previous frame, a finite search window is set in the current frame. The size of the search window can be determined according to the expected maximum pixel offset. Within the search window, the cross-correlation algorithm is used to find the region most similar to the image patch around the feature point in 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 two image patches. The higher the correlation, the higher the similarity. By comparing the positions of the feature point in the previous frame and the current frame, the pixel offset of the feature point between adjacent frames is calculated, and these offsets reflect the movement of the image in the local area.

[0064] More specifically, thus, in step S23, using the pixel offsets of multiple calculated feature points, a global affine transformation matrix is fitted. Affine transformation is a linear geometric transformation that can describe motions such as translation, rotation, scaling, and shearing of an image. Through robust fitting methods such as the least squares method 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 the small jitters existing in the image sequence, ensuring that subsequent processing is based on a relatively stable image perspective. For example, for small translations and rotations, the affine transformation can provide an accurate description.

[0065] More specifically, in step S24, bilinear interpolation is a commonly used image resampling technique. It calculates the gray value of the point to be interpolated based on the gray values of four surrounding known pixel points and their distances from the point to be interpolated. Using bilinear interpolation for position correction can smoothly transform the image, avoid jagged edges, and generate the corrected image. The corrected images are combined to form a corrected image sequence and provided to subsequent steps for feature extraction and target detection. Through this precise position correction, the influence caused by small jitters can be effectively eliminated, making subsequent feature extraction more stable and target detection more accurate.

[0066] More specifically, through the above technical solution, the present application solves the problem of frame - to - frame misalignment in the image sequence caused by small jitters during image acquisition. Through feature - point - based matching and global affine transformation, the small displacements of the image are accurately calculated and compensated. This position - correction process generates a corrected image sequence with frame - to - 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 blurring and position uncertainty caused by misalignment, thereby supporting the recognition of contaminated targets.

[0067] In some preferred embodiments, step S5 includes: S51. For the high - resolution feature map and the low - resolution feature map, weight parameter matrices corresponding to the corrected image are respectively constructed according to the weight parameter group; S52. According to the weight parameter matrix corresponding to the high - resolution feature map, the high - resolution weight parameters of the image blocks corresponding to each pixel point in the high - resolution feature map are extracted, and the feature vectors of the corresponding pixel points in the high - resolution feature map are adjusted according to the high - resolution weight parameters to obtain the adjusted high - resolution feature map; S53. Extract the low-resolution weight parameters of the image patches corresponding to each pixel in the low-resolution feature map according to the weight parameter matrix corresponding to the low-resolution feature map, and adjust the feature vectors of the corresponding pixels in the low-resolution feature map according to the low-resolution weight parameters 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.

[0068] Specifically, in step S51, the weight parameter group includes weight parameters calculated based on multiple image patches divided from the corrected image, and each image patch corresponds to one or a group of weight parameters.

[0069] 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 patch to all pixel positions corresponding to the image patch on the feature map, or by upsampling (e.g., using bilinear interpolation) the weight parameters of the image patch to the resolution of the feature map to construct the weight parameter matrix. Steps S52 and S53 associate the local characteristic weights in 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.

[0070] More specifically, step S54 combines two different types of feature information that have both been optimized according to local characteristics by concatenating the high- and low-resolution feature maps adjusted by the weight parameter matrix in the channel dimension. This fusion method enables the final feature map to contain both high-frequency details that have been locally enhanced or suppressed and robust low-frequency information. This step, combined with the step of analyzing and generating the weight parameter group according to local region characteristics, realizes the weighted adjustment of the feature map based on the local characteristics of the image. This adjustment is performed before feature fusion, enabling the fusion process to utilize the local information of the image to guide the contributions of different-resolution features, thereby generating a fused feature map with a richer, more discriminative, and more locally adaptable feature representation, thus improving the detection accuracy for micro-pollution targets.

[0071] In some preferred embodiments, step S3 includes: S31. Preprocess the corrected image to be recognized, and the preprocessing includes image denoising, contrast enhancement, and normalization operations to obtain a preprocessed image. S32. Input the preprocessed image into a preset high-resolution convolutional neural network to extract a shallow high-resolution feature map. The high-resolution convolutional neural network includes multiple convolutional layers and pooling layers for retaining the high-frequency detail information of the image. 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, and is used to extract robust low-frequency features of the image.

[0072] Specifically, step S31 preprocesses the corrected image to be recognized, aiming to improve the image quality and provide a better input for subsequent feature extraction.

[0073] More specifically, steps S32 and S33 extract features from the preprocessed image in parallel.

[0074] More specifically, step S32 inputs the preprocessed image into a preset high-resolution convolutional neural network. The design of this network focuses on retaining the high-frequency detail information of the image, so it includes multiple convolutional layers and pooling layers. The convolutional layer extracts local features of the image through convolutional kernels, and the pooling layer retains the main features while reducing the spatial size of the feature map. The shallow feature map extracted by this network usually has a high spatial resolution and contains rich high-frequency details such as edges and textures, which is crucial for identifying tiny pollution targets that rely on fine structures.

[0075] 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 convolutional layers and pooling layers, it also includes downsampling layers. 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 context information of the image. Through multiple layers of convolution and downsampling, the deep feature map extracted by this network has a low spatial resolution, but has better robustness to translation, rotation, and scale changes of the image, that is, it extracts the 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 the resistance to noise and background changes.

[0076] Thus, the features output by step S3 combine the feature representations of detail retention and robust extraction, overcoming the limitation that a single processing path is difficult to adapt to complex imaging conditions, providing more effective information for subsequent feature fusion and target detection, and thus improving the performance of identifying tiny pollution targets in complex and changing environments.

[0077] In some preferred embodiments, step S6 includes: S61. Perform a convolution operation on the fused feature map with each convolution kernel in a preset set of convolution kernels to obtain multiple response maps corresponding to the pollution types. Each convolution kernel in the set of convolution kernels enhances the response to different types of pollution targets; S62. For each response map, the non-maximum suppression algorithm is used for post-processing to screen out the candidate boxes of pollution targets with high confidence. According to the positions, sizes and response intensities of the candidate boxes, combined with the mapping relationship between pollution types and response intensities, the pollution recognition results are output.

[0078] Specifically, in step S61, the fused feature map combines high-resolution details and robust low-frequency features, containing rich target information. By convolving with a preset convolution kernel for response enhancement for specific pollution types, stronger responses can be generated in the regions related to specific pollution types in the fused feature map, thus highlighting pollution targets of different types, suppressing background interference, and laying a foundation for subsequent accurate detection. The design of each convolution kernel can capture the pattern features of specific pollution types, enabling the convolution operation to effectively convert the abstract features in the fused feature map into spatial response intensities related to specific pollution types.

[0079] More specifically, in step S62, by applying non-maximum suppression to the response map, candidate boxes representing potential pollution targets can be extracted from the response map, and it is ensured that each target is detected only once, improving the neatness and accuracy of the detection results. According to the positions, sizes and response intensities of the selected candidate boxes, combined with the preset mapping relationship between pollution types and response intensities, the final pollution recognition results are output.

[0080] More specifically, through the above technical solution, this application solves the problem that it is difficult to accurately identify tiny pollution targets directly based on the fused feature map under complex dynamic imaging conditions. By convolving the fused feature map with a convolution kernel preset for specific pollution types, the response of tiny pollution targets in the feature map can be effectively enhanced, while suppressing the interference of background noise and artifacts, making pollution targets of different types stand out.

[0081] In a second aspect, please refer to Figure 2 , some embodiments of this application further provide a polypeptide drug membrane filtration pollution recognition system for a convolutional neural network. The system includes: An acquisition module 201, configured to acquire the image to be recognized on the surface of the filter membrane and the corresponding image sequence; A correction module 202, configured to perform position correction on the image sequence to generate a corrected image sequence including the corrected image to be recognized; A processing module 203, configured to input the corrected image to be recognized in the corrected image sequence into a preset high-resolution processing path and a low-resolution processing path to obtain a high-resolution feature map and a low-resolution feature map; A weight module 204, configured to analyze the local region characteristics of the corrected image to be recognized to generate a weight parameter group; A fusion module 205, configured to generate a fused feature map by fusing the high-resolution feature map and the low-resolution feature map according to the weight parameter group; An identification module 206, configured to perform object detection based on the fused feature map to generate a pollution identification result.

[0082] The core innovation of the polypeptide drug membrane filtration pollution identification system of the convolutional neural network according to the embodiments of the present application lies in combining high- and low-resolution dual-path processing and weighted fusion based on local region characteristics, so as to generate a fused feature map for object detection, effectively addressing 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 miss rate.

[0083] In addition, the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to 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.

[0084] Furthermore, in each embodiment of the present application, the various functional modules may be integrated together to form an independent part, or each module may exist alone, or two or more modules may be integrated to form an independent part.

[0085] In this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations.

[0086] The above are only the embodiments of the present application and are not intended to limit the protection scope of the present application. For those skilled in the art, the present application may have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for identifying polypeptide drug membrane filtration pollution in a convolutional neural network, characterized in that, The method includes the following steps: S1. Obtain the image to be recognized on the surface of the filter membrane and the corresponding image sequence; S2. Perform position correction on the image sequence to generate a corrected image sequence including the corrected image to be recognized; S3. Input the corrected image to be recognized in the corrected image sequence into a preset high-resolution processing path and a low-resolution processing path to obtain a high-resolution feature map and a low-resolution feature map; S4. Analyze the local region characteristics of the corrected image to be recognized 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. Perform target detection based on the fused feature map to generate a pollution recognition result; The local region characteristics include local contrast, local texture complexity, and local motion degree. Step S4 includes: S41. Calculate the gradient magnitude and gradient direction of each pixel point in the corrected image to be recognized; S42. According to the gradient magnitude and the gradient direction, divide the corrected image to be recognized into multiple image blocks, and calculate the local contrast, local texture complexity, and local motion degree of each image block; S43. According to the local contrast, local texture complexity, and local motion degree of each image block, use a preset non-linear mapping model to calculate the initial weight parameters of each image block corresponding to the high-resolution processing path and the low-resolution processing path; S44. Perform normalization processing on the initial weight parameters to obtain the final weight parameter group.

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

3. A method for identifying polypeptide drug membrane filtration pollution of 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. Calculate the standard deviation of all pixel gray values in each image block and divide it by the mean of all pixel gray values in the image block to obtain the normalized local contrast; S424. Calculate the histogram of the gradient directions of all pixel points in each image block, count the number of pixel points in each gradient direction, and calculate the entropy value of the histogram as the local texture complexity; S425. Calculate the optical flow vector between each image block and the corresponding image block in the previous frame, and count the average magnitude of the optical flow vector to obtain the local motion degree.

4. A method for identifying polypeptide drug membrane filtration pollution in a convolutional neural network according to claim 1, characterized in that, The step of dividing the corrected image to be recognized into multiple image blocks according to the gradient magnitude and the gradient direction includes: S421. Set the weight of the pixel point according to the gradient magnitude. The larger the gradient magnitude, the higher the weight; S422. Use the weighted K-means clustering algorithm, with the position coordinates of the pixel points as features and the weight of the pixel points as the clustering weight, to divide the corrected image to be recognized into multiple image blocks.

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

6. A method for identifying polypeptide drug membrane filtration pollution of 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, construct weight parameter matrices corresponding to the corrected image respectively according to the weight parameter group; S52. According to the weight parameter matrix corresponding to the high-resolution feature map, extract the high-resolution weight parameters of the image blocks corresponding to each pixel point in the high-resolution feature map, and adjust the feature vectors of the corresponding pixel points in the high-resolution feature map according to the high-resolution weight parameters to obtain an adjusted high-resolution feature map; S53. According to the weight parameter matrix corresponding to the low-resolution feature map, extract the low-resolution weight parameters of the image blocks corresponding to each pixel point in the low-resolution feature map, and adjust the feature vectors of the corresponding pixel points in the low-resolution feature map according to the low-resolution weight parameters 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. A method for identifying polypeptide drug membrane filtration pollution in a convolutional neural network according to claim 1, characterized in that, Step S3 includes: S31. Preprocess the corrected image to be recognized to obtain a preprocessed image; S32. Input the preprocessed image into a preset high-resolution convolutional neural network to extract a high-resolution feature map. 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 convolutional layers, pooling layers and downsampling layers.

8. A method for identifying polypeptide drug membrane filtration pollution of a convolutional neural network according to claim 1, characterized in that, Step S6 includes: S61. Perform convolution operations on the fused feature map with each convolution kernel in a preset convolution kernel set to obtain multiple response maps corresponding to the pollution types. Each convolution kernel in the convolution kernel set enhances the response to different types of pollution targets; S62. For each response map, perform post-processing using the non-maximum suppression algorithm to screen the candidate boxes of pollution targets with high confidence. According to the positions, sizes and response intensities of the candidate boxes, combined with the mapping relationship between the pollution type and the response intensity, output the pollution recognition result.

9. A polypeptide drug membrane filtration pollution recognition system for a convolutional neural network, characterized in that, The system includes: An acquisition module for acquiring the image to be recognized on the surface of the filter membrane and the corresponding image sequence; A correction module for performing position correction on the image sequence to generate a corrected image sequence including the corrected image to be recognized; A processing module for inputting the corrected image to be recognized in the corrected image sequence into a preset high-resolution processing path and a low-resolution processing path to obtain a high-resolution feature map and a low-resolution feature map; A weight module for analyzing the local region characteristics of the corrected image to be recognized to generate a weight parameter group; A fusion module, configured to fuse 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 region characteristics include local contrast, local texture complexity, and local motion degree. The process of analyzing the local region characteristics of the image to be identified after correction to generate a weight parameter group includes: S41. Calculate the gradient magnitude and gradient direction of each pixel point in the image to be identified after correction; S42. According to the gradient magnitude and the gradient direction, divide the image to be identified after correction into a plurality of image blocks, and calculate the local contrast, local texture complexity, and local motion degree of each image block; S43. According to the local contrast, local texture complexity, and local motion degree of each image block, use a preset non-linear mapping model to calculate the initial weight parameters corresponding to the high-resolution processing path and the low-resolution processing path of each image block; S44. Perform normalization processing on the initial weight parameters to obtain the final weight parameter group.

Images