Image processing method and device for low-power-consumption MCU (Microprogrammed Control Unit) visual module

Abstract

The invention relates to the technical field of image data processing, and provides an image processing method and device for a low-power-consumption MCU visual module, and the method comprises the steps: obtaining original image data collected by a sensor, carrying out the de-noising processing, carrying out the target recognition after obtaining clear image data, obtaining a target region, inputting the target region into a preset deep learning model, and obtaining a deep learning model; the method comprises the following steps: obtaining feature data, carrying out dimension reduction processing to obtain a feature vector after dimension reduction so as to identify a plurality of items of category data and category confidence data of a target area, reconstructing original image data to obtain processed image data, inputting the processed image data into an image compression algorithm for calculation, and obtaining an image compression result. And obtaining compressed image data. By optimizing reconstruction and compression of a target area, the effect of efficiently processing complex image data in a low-power-consumption environment is achieved, and the problems that when a complex visual task is processed, the classification precision is low, and complex features cannot be effectively extracted are solved.

Description

Technical Field

[0001] The present application relates to the technical field of image data processing, and in particular to an image processing method and device for a low-power MCU vision module. Background Art

[0002] With the rapid development of Internet of Things (IoT) technology and the widespread popularity of smart devices, the demand for low-power MCU (Microcontroller Unit) vision modules in edge computing has increased significantly. These modules play an important role in scenarios such as smart homes, industrial monitoring, and medical equipment. In order to achieve efficient processing of image data while meeting the requirements of low power consumption and cost, the research and development of related technologies has become the focus of industry attention.

[0003] In related technical means, the MCU vision module performs image processing, usually by acquiring raw image data through sensors, and using basic image preprocessing techniques (such as filtering, edge detection, and simple threshold segmentation) to process the image. Subsequently, the target area is identified through a lightweight target detection algorithm or a rule-based method. These methods can achieve a certain degree of target detection and classification tasks to meet low-complexity application requirements, such as simple environmental monitoring or identification of specific objects.

[0004] Regarding the above technical solutions, although the basic image processing functions of low-power devices can be realized through these traditional image preprocessing methods and lightweight target detection algorithms, when processing complex visual tasks, such as the need to accurately identify multi-target scenes or classify fine-grained target categories, there are problems such as low classification accuracy and inability to effectively extract complex features. Summary of the invention

[0005] In order to improve the problems of low classification accuracy and inability to effectively extract complex features when processing complex visual tasks, the present application provides an image processing method and device for a low-power MCU vision module.

[0006] The present invention provides an image processing method for a low-power MCU vision module, comprising: acquiring raw image data collected by a sensor, and performing denoising processing on the raw image data to obtain clear image data; performing target recognition based on the clear image data to obtain a target area, inputting the target area into a preset deep learning model to obtain feature data, performing dimensionality reduction processing on the feature data to obtain a feature vector after dimensionality reduction; identifying several items of category data and category confidence data of the target area according to the feature vector after dimensionality reduction, performing clustering processing on all the category data to generate a clustering result, optimizing the clustering result using the category confidence data to obtain an optimized clustering result, optimizing the target area according to the optimized clustering result to obtain an optimized target area; reconstructing the raw image data using the optimized target area to obtain processed image data, and inputting the processed image data into an image compression algorithm for calculation to obtain compressed image data.

[0007] As a preferred solution, the step of acquiring the original image data collected by the sensor and denoising the original image data to obtain clear image data includes: collecting the original image data output by the sensor and the noise feature data corresponding to the original image data, denoising the original image data and the noise feature data using a multi-scale filtering method to obtain preliminary denoised image data and residual noise data; constructing a noise mapping relationship through the residual noise data, and performing adaptive noise reduction processing on the preliminary denoised image data based on the noise mapping relationship to obtain target denoised image data; performing artifact removal on the target denoised image data to obtain optimized image data and artifact feature data, calculating the gradient change information of the optimized image data, and performing edge refinement processing on the gradient change information and the artifact feature data to obtain clear image data.

[0008] As a preferred solution, the steps of removing artifacts from the target denoised image data to obtain optimized image data and artifact feature data, calculating gradient change information of the optimized image data, performing edge refinement processing on the gradient change information and the artifact feature data to obtain clear image data include: performing spectral decomposition on the target denoised image data based on a frequency domain analysis method to obtain component data, extracting artifact feature data based on the component data, performing artifact suppression on the component data based on the artifact feature data using a directional filtering method to obtain artifact suppressed image data and artifact residual data; calculating local gradient change information of the artifact suppressed image data, and adaptively compensating the artifact residual data based on the local gradient change information to obtain optimized image data; applying an interpolation method to reconstruct edges of the optimized image data to obtain edge optimized image data and edge error information, calculating overall texture information of the edge optimized image data, and performing texture enhancement processing on the edge optimized image data based on the overall texture information and edge error information to obtain clear image data.

[0009] As a preferred solution, the step of performing target recognition based on the clear image data to obtain a target area, inputting the target area into a preset deep learning model to obtain feature data, performing dimensionality reduction processing on the feature data to obtain a feature vector after dimensionality reduction includes: performing feature stratification on the clear image data to obtain global feature data and local feature data, applying a domain adaptive image segmentation method to perform region extraction on the global feature data and the local feature data to obtain multiple candidate target areas and regional confidence data; screening multiple candidate target areas based on the regional confidence data using a preset feature fusion strategy to obtain a target area and target feature data; inputting the target area and the target feature data into a preset deep learning model, applying a feature mapping method of a Transformer self-attention mechanism in the deep learning model to perform feature mapping calculation on the target area and the target feature data to obtain an original feature vector and spatial feature relationship data; based on the spatial feature relationship data, using a graph analysis method to perform feature association on the original feature vector to obtain an associated feature vector and feature importance data, and performing dimensionality reduction calculation on the original feature vector using the associated feature vector and feature importance data to obtain a feature vector after dimensionality reduction.

[0010] As a preferred solution, the steps of identifying several items of category data and category confidence data of the target area according to the feature vector after dimension reduction, clustering all the category data to generate clustering results, optimizing the clustering results using the category confidence data to obtain optimized clustering results, optimizing the target area according to the optimized clustering results to obtain the optimized target area include: inputting the feature vector after dimension reduction into a preset target area feature model, applying a multi-classification method in the target area feature model to classify the target area based on the feature vector after dimension reduction to obtain several items of category data and category confidence data Category confidence data; hierarchically clustering all the category data to obtain clustering results and category center data, optimizing and adjusting the clustering results using the category confidence data and the category center data to obtain optimized clustering results; calculating the feature distance between the feature vector after dimensionality reduction and the optimized clustering result to obtain matching error data, adaptively correcting the optimized clustering result according to the matching error data to obtain corrected category data; identifying the category distribution information of the target area based on the corrected category data, inputting the category distribution information into the target area feature model for category optimization processing, and obtaining the optimized target area.

[0011] As a preferred scheme, the step of calculating the feature distance between the feature vector after dimensionality reduction and the optimized clustering result to obtain matching error data, and adaptively correcting the optimized clustering result according to the matching error data to obtain corrected category data includes: calculating the Euclidean distance, cosine similarity and Mahalanobis distance between the feature vector after dimensionality reduction and the optimized clustering result, calculating the matching error data based on the Euclidean distance and the Mahalanobis distance, and generating feature similarity data using the cosine similarity; applying a dynamic weight adjustment method to normalize the matching error data based on the feature similarity data to obtain normalized matching error data and feature contribution data, performing error compensation calculation on the normalized matching error data using the feature contribution data to obtain optimized matching error data and error distribution information; adaptively adjusting the optimized matching error data using the error distribution information to obtain error correction parameters, and performing category center migration adjustment on the optimized clustering result according to the error correction parameters to obtain corrected category data.

[0012] As a preferred scheme, the steps of reconstructing the original image data using the optimized target area to obtain processed image data, and inputting the processed image data into an image compression algorithm for calculation to obtain compressed image data include: comparing the optimized target area with other areas in the original image data to obtain a comparison result, reconstructing the target area in the original image data based on the comparison result by an image reconstruction algorithm to obtain a reconstructed target area; fusing the reconstructed target area with other areas in the original image data to obtain fused image data, preprocessing the fused image data to obtain processed image data, and compressing the processed image data using an image compression algorithm to obtain compressed image data.

[0013] The present application also provides an image processing device of a low-power MCU vision module, including: an acquisition module, used to acquire raw image data collected by a sensor, and perform denoising on the raw image data to obtain clear image data; a dimension reduction module, used to perform target recognition based on the clear image data to obtain a target area, input the target area into a preset deep learning model to obtain feature data, perform dimension reduction on the feature data to obtain a feature vector after dimension reduction; a clustering module, used to identify several items of category data and category confidence data of the target area according to the feature vector after dimension reduction, perform clustering on all the category data to generate a clustering result, optimize the clustering result using the category confidence data to obtain an optimized clustering result, optimize the target area according to the optimized clustering result to obtain an optimized target area; a compression module, used to reconstruct the raw image data using the optimized target area to obtain processed image data, input the processed image data into an image compression algorithm for calculation to obtain compressed image data.

[0014] Compared with the prior art, the present application has the following beneficial effects: high classification accuracy and strong recognition features. By using median filtering and adaptive filtering, the noise of the original image data is efficiently removed, ensuring the quality of clear image data; the feature data of the target area is extracted through a lightweight deep learning model, and the subsequent calculation complexity is simplified by combining dimensionality reduction processing; the classification accuracy of the target area and the matching degree of the target boundary are significantly improved through clustering algorithms and confidence optimization; by optimizing the reconstruction and compression of the target area, the effect of efficient processing of complex image data in a low-power environment is achieved, which effectively solves the low classification accuracy and data redundancy of the low-power MCU visual module in complex visual tasks, and improves the problem of low classification accuracy and inability to effectively extract complex features when processing complex visual tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0016] The structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantive technical significance. Any structural modification, change in proportion or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical contents disclosed by the present invention.

[0017] Figure 1 It is a flowchart of an image processing method of a low-power MCU vision module provided by an embodiment of the present invention; Figure 2 It is a schematic block diagram of the structure of an image processing device of a low-power MCU vision module provided in an embodiment of the present invention.

[0018] Description of reference numerals: 10. Image processing device of low-power MCU vision module; 11. Acquisition module; 12. Dimensionality reduction module; 13. Clustering module; 14. Compression module. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0021] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0022] It should be further understood that the term “and / or” used in the specification and appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0023] The technical solution of the present invention is further described below with reference to the accompanying drawings and through specific implementation methods.

[0024] Embodiment 1: like Figure 1 As shown, the present application provides an image processing method for a low-power MCU vision module, including steps S100 to S400.

[0025] Step S100: acquiring original image data collected by the sensor, and performing denoising processing on the original image data to obtain clear image data.

[0026] In this step, after acquiring the original image data through the sensor, the image denoising algorithms such as median filtering, mean filtering and adaptive filtering are used to suppress the noise signal in the original image data, eliminate high-frequency noise and irregular interference signals, and obtain clear image data with a clean background and clear edges. Specifically, combined with the limitations of hardware resources, a low-complexity filtering method is selected according to the noise characteristics of the original image data, and multi-window filtering technology is used to further enhance the image detail expression.

[0027] For example, in monitoring equipment in industrial environments, denoising can effectively filter out high-frequency noise caused by vibration or light changes, ensuring the quality of the original image data and providing a better input basis for subsequent processing.

[0028] Step S200: perform target recognition based on clear image data to obtain a target area, input the target area into a preset deep learning model to obtain feature data, perform dimensionality reduction processing on the feature data, and obtain a feature vector after dimensionality reduction.

[0029] In this step, the target area in the clear image data is accurately located by adopting a target recognition algorithm based on edge detection and region growing, and the features of the target area are extracted using a deep learning model (such as a convolutional neural network). Specifically, a lightweight network structure (such as MobileNet) is used in the deep learning model to adapt to the computing power limitations of the low-power MCU, and the extracted feature data is reduced in dimension by the principal component analysis (PCA) method to remove redundant features and obtain a more discriminative feature vector after dimensionality reduction.

[0030] For example, in smart camera applications in smart homes, the deep learning model can accurately extract the key feature points of the facial area, and improve data processing efficiency and device response speed through dimensionality reduction processing.

[0031] Step S300: Identify several items of category data and category confidence data of the target area based on the feature vector after dimensionality reduction, cluster all the category data to generate clustering results, optimize the clustering results using the category confidence data to obtain optimized clustering results, optimize the target area based on the optimized clustering results to obtain an optimized target area.

[0032] In this step, the feature vector after dimensionality reduction is input into a clustering algorithm (such as the K-Means algorithm) for category clustering processing, and the clustering result is combined with the category confidence data for confidence-weighted optimization processing to make the clustering result more accurate; specifically, the optimized clustering result is used to further adjust the boundary of the target area to ensure that the optimized target area matches the actual target more closely.

[0033] For example, in microscope image analysis in medical equipment, the target cell area can be located through clustering results of category data, and its boundaries can be optimized to more accurately identify specific cell categories.

[0034] Step S400: reconstructing the original image data using the optimized target area to obtain processed image data, and inputting the processed image data into an image compression algorithm for calculation to obtain compressed image data.

[0035] In this step, the local features in the original image data are enhanced and reconstructed through the optimized target area, and the processed image data is efficiently compressed using image compression algorithms such as discrete cosine transform (DCT) or discrete wavelet transform (DWT) to obtain compressed image data; specifically, the compression coefficient is dynamically adjusted to ensure that the compression rate is maximized and key image information is retained in a low power environment.

[0036] For example, when transmitting surveillance images on edge devices, compressed image data can significantly reduce bandwidth usage while maintaining the clarity of the target area to meet real-time transmission requirements.

[0037] In this embodiment, by acquiring the original image data collected by the sensor and performing denoising on the original image data, clear image data is obtained. Target recognition is performed based on the clear image data, the target area is extracted, and the target area is input into a preset deep learning model to generate feature data. Subsequently, the feature data is subjected to dimensionality reduction processing to obtain a feature vector after dimensionality reduction. According to the feature vector after dimensionality reduction, several items of category data and category confidence data of the target area are identified, and all category data are clustered to generate clustering results. Then, the clustering results are optimized by using the category confidence data to obtain an optimized clustering result, and the target area is optimized by using the optimized clustering result to obtain an optimized target area. Finally, the original image data is reconstructed using the optimized target area to generate processed image data, and the processed image data is input into the image compression algorithm for calculation to obtain compressed image data.

[0038] The quality of the original image data is improved through denoising, laying the foundation for subsequent processing; the deep learning model combined with dimensionality reduction processing reduces the complexity of the feature data while retaining key information, improving the efficiency and accuracy of target recognition; clustering processing and optimization steps further enhance the accuracy of classification and reduce the interference of redundant data; finally, through the reconstruction and compression steps, not only the amount of data in the storage and transmission process can be significantly reduced, but the main features of the processed image data can be retained, thereby achieving efficient processing of complex visual tasks under limited resources and power consumption, and improving the problems of low classification accuracy and inability to effectively extract complex features when processing complex visual tasks.

[0039] Embodiment 2: In step S100, the original image data output by the sensor and the noise characteristic data corresponding to the original image data are collected, and the original image data and the noise characteristic data are denoised using a multi-scale filtering method to obtain image data after preliminary denoising and residual noise data.

[0040] Through the multi-scale filtering method, combined with the spectral characteristics of the original image data and noise feature data output by the sensor, adaptive filters in multiple frequency bands are designed to perform targeted processing on low-frequency, medium-frequency and high-frequency signals respectively; specifically, a low-pass filter is used to filter out low-frequency noise, a band-pass filter is used to suppress interference in the medium-frequency signal, and a high-pass filter is used to weaken the sharp noise in the high-frequency signal, so as to achieve layer-by-layer decomposition and reconstruction of the original image data, so as to retain the image detail information to the maximum extent and suppress noise interference.

[0041] For example, for industrial equipment monitoring scenarios, periodic noise caused by mechanical vibration in the original image data can be effectively removed through multi-scale filtering methods, while retaining the clear texture details of the gear or bearing surface, providing high-quality data for subsequent image analysis.

[0042] A noise mapping relationship is constructed through the residual noise data, and adaptive denoising is performed on the image data after preliminary denoising based on the noise mapping relationship to obtain the target denoised image data.

[0043] By analyzing the spatial distribution characteristics of the residual noise data, a noise mapping matrix is ​​constructed, and the denoising parameters are dynamically adjusted based on the mapping matrix to make the denoising process more in line with the actual needs of different image regions. Specifically, by applying a stronger filtering strategy to high-noise areas and taking a weaker denoising operation to low-noise areas, it is ensured that image details are retained while reducing noise.

[0044] For example, in surgical microscope image analysis, low-noise data in vascular areas can retain more details, while high noise in background areas can be more strongly reduced, ultimately improving the overall image clarity and diagnostic reliability.

[0045] The target denoised image data is subjected to artifact removal to obtain optimized image data and artifact feature data, the gradient change information of the optimized image data is calculated, and the gradient change information and artifact feature data are subjected to edge refinement processing to obtain clear image data.

[0046] By performing frequency domain analysis on the target denoised image data, the high-frequency components that produce artifacts are separated and the artifact feature data is extracted. The directional filtering algorithm is used to perform targeted processing on the artifact components to eliminate artifacts at the image edges and complex textures. Specifically, the component decomposition method is used to locally separate the artifacts, and then multi-scale suppression is implemented on the artifact residual data. The edge refinement processing is enhanced by combining gradient information to ensure the continuity and realism of the image edges.

[0047] For example, in remote sensing satellite image processing, by eliminating artifacts at the boundaries of clouds or objects, the optimized images can be used for accurate landform analysis and resource assessment.

[0048] Among them, the steps of removing artifacts from the target denoised image data to obtain optimized image data and artifact feature data, calculating the gradient change information of the optimized image data, and performing edge refinement processing on the gradient change information and the artifact feature data to obtain clear image data include: performing spectral decomposition on the target denoised image data based on a frequency domain analysis method to obtain component data, and extracting artifact feature data based on the component data, and suppressing artifacts on the component data based on the artifact feature data using a directional filtering method to obtain image data and artifact residual data after artifact suppression.

[0049] By performing FFT (Fast Fourier Transform) on the target denoised image data, the artifact components in the spectrum are separated, and the artifact distribution is located using the artifact feature extraction algorithm. Specifically, the filtering direction is adjusted in combination with the artifact feature data to dynamically eliminate high-frequency artifact residues, ensuring effective suppression of artifacts without damaging image details.

[0050] For example, in the processing of medical images (such as X-rays), removing artifacts can significantly improve the clarity of tissue boundaries, thereby improving the doctor's diagnostic accuracy of lesions.

[0051] The local gradient change information of the image data after artifact suppression is calculated, and the artifact residual data is adaptively compensated based on the local gradient change information to obtain optimized image data.

[0052] By calculating the local gradient change information of the image data after artifact suppression, the gradient amplitude and directional characteristics of the image edge area are extracted, and a gradient change mapping model is constructed. Using this model, the distribution of the artifact residual data on the local gradient characteristics is analyzed, and the compensation parameters are adjusted through adaptive weight allocation, and the artifact residual data is directional suppressed and intensity optimized, and finally the optimized image data is generated; specifically, an enhanced filtering algorithm is used in the high gradient change area to strengthen the retention of edge features, while reducing the interference of the artifact residual in the low gradient change area, ensuring the improvement of the overall image quality.

[0053] For example, in medical images such as MRI scans, accurate compensation of artifacts at the edge of a tumor can improve the clarity of the edge contour, thus providing a more reliable basis for clinical diagnosis.

[0054] The interpolation method is applied to reconstruct the edges of the optimized image data to obtain the edge-optimized image data and edge error information, the overall texture information of the edge-optimized image data is calculated, and the texture enhancement processing is performed on the edge-optimized image data based on the overall texture information and edge error information to obtain clear image data.

[0055] By applying interpolation methods (such as bicubic interpolation or adaptive interpolation), the pixel values ​​of the edge areas in the optimized image data are reconstructed, the areas with discontinuous or blurred edges are filled, the edge-optimized image data is generated, and the error information in the interpolation process is recorded as edge error information; specifically, the texture features are extracted in combination with the overall texture analysis algorithm (such as local binary pattern LBP), and the performance of the edge-optimized image data in local details and overall smoothness is improved through texture enhancement technology, ensuring that clear image data with high visual quality is obtained.

[0056] For example, in smart security, edge reconstruction and texture enhancement of images collected in low-light environments can effectively improve the clarity and contrast of facial areas, thereby improving the accuracy of target recognition.

[0057] In step S200, the clear image data is feature-stratified to obtain global feature data and local feature data, and a domain adaptive image segmentation method is applied to perform region extraction on the global feature data and the local feature data to obtain multiple candidate target regions and region confidence data.

[0058] By performing hierarchical analysis on clear image data, the overall geometric features of the image are extracted as global feature data, while the detailed texture and shape characteristics are extracted as local feature data. Using the domain adaptive image segmentation method, adaptive segmentation strategies are applied to global feature data and local feature data respectively, combined with edge detection and region growing algorithms, multiple candidate target regions in the image are automatically extracted, and regional confidence data is calculated for each region to measure its target relevance; specifically, by dynamically adjusting the segmentation parameters, the segmentation results are adapted to the image characteristics in different scenarios.

[0059] For example, in autonomous driving scenarios, feature stratification and domain adaptive segmentation can effectively extract candidate target areas of road boundaries, vehicles, and pedestrians, and provide highly reliable data support for perception tasks in complex environments.

[0060] Based on the regional confidence data, a preset feature fusion strategy is used to screen multiple candidate target regions to obtain the target region and target feature data.

[0061] By analyzing the regional confidence data of multiple candidate target areas, a preset feature fusion strategy is used to comprehensively consider the geometric shape, texture features and position distribution of the target area to screen the target area; specifically, a multi-layer perception mechanism is used to screen the area with the highest confidence as the target area, and at the same time, key features are further extracted from the target area to generate target feature data.

[0062] For example, in drone monitoring, filtering target areas of interest (such as a specific building or vehicle) through fusion strategies can improve the efficiency of target recognition and provide key input for subsequent in-depth analysis.

[0063] The target area and target feature data are input into a preset deep learning model, and the feature mapping method of the Transformer self-attention mechanism is applied in the deep learning model to perform feature mapping calculations on the target area and target feature data to obtain the original feature vector and spatial feature relationship data.

[0064] By inputting the target area and target feature data into a preset deep learning model (such as a Transformer-based structure), the dynamic weight allocation method of the self-attention mechanism is used to perform feature mapping and relationship modeling on each pixel point in the target area to generate an original feature vector. At the same time, spatial feature relationship modeling is used to further capture the interactive relationship between the internal features of the target area and generate spatial feature relationship data. Specifically, this process reduces the computational complexity of the model while ensuring feature accuracy.

[0065] For example, in high-resolution remote sensing images, feature mapping and relationship modeling can accurately capture the geometric shape and internal characteristics of farmland or built-up areas, providing support for precise mapping.

[0066] Based on the spatial feature relationship data, the graph analysis method is used to perform feature association on the original feature vector to obtain the associated feature vector and feature importance data. The associated feature vector and feature importance data are used to perform dimensionality reduction calculation on the original feature vector to obtain the feature vector after dimensionality reduction.

[0067] By constructing a feature graph based on spatial feature relationship data, the original feature vectors are used as graph nodes, and the feature associations are calculated through the relationship weight values ​​between nodes to obtain the associated feature vectors and feature importance data. Specifically, the feature graph is reduced in dimension by combining a dimensionality reduction algorithm (such as principal component analysis or graph embedding method) to ensure that the feature vectors after dimensionality reduction reduce data redundancy while retaining the main features.

[0068] For example, in low-power smart camera devices, feature dimensionality reduction processing can reduce the computing burden of the device and improve the efficiency of detecting target features in dynamic scenes.

[0069] In step S300, the reduced-dimensional feature vector is input into a preset target region feature model, in which a multi-classification method is applied to classify the target region based on the reduced-dimensional feature vector to obtain a number of category data and category confidence data.

[0070] By inputting the reduced-dimensional feature vector into the target region feature model, the model performs multi-classification processing on the target region based on a neural network structure (such as a Softmax classifier). Specifically, by constructing a multi-layer classification network, the reduced-dimensional feature vector is subjected to layer-by-layer feature mapping and classification scoring, and finally outputs several items of category data and corresponding category confidence data, where the category confidence data indicates the credibility of the classification result.

[0071] For example, in an intelligent transportation system, by classifying the feature vectors of target vehicles, the type of vehicle (such as small cars, large cars, motorcycles, etc.) and the confidence of the classification can be identified, providing high-precision support for traffic flow analysis.

[0072] All category data are hierarchically clustered to obtain clustering results and category center data. The clustering results are optimized and adjusted using category confidence data and category center data to obtain optimized clustering results.

[0073] By performing hierarchical clustering on several category data, the bottom-up clustering method is used to gradually merge categories to generate a cluster tree structure, and the clustering results are optimized according to the Euclidean distance of the category center data and the category confidence data, and the accuracy of the clustering level and category division is adjusted. Specifically, the cutoff point of the clustering result is fine-tuned in combination with the geometric distribution characteristics of the category center and the weight of the category confidence data to improve the rationality and consistency of the clustering.

[0074] For example, in biological microscopic image processing, multiple similar cell categories can be integrated through clustering analysis and optimization to generate clustering results that are more in line with actual biological significance for subsequent pathological analysis.

[0075] The characteristic distance between the feature vector after dimensionality reduction and the optimized clustering result is calculated to obtain matching error data, and the optimized clustering result is adaptively corrected according to the matching error data to obtain corrected category data.

[0076] By calculating the Euclidean distance, cosine similarity and Mahalanobis distance between the feature vector after dimensionality reduction and the optimized clustering result, the matching degree is comprehensively analyzed to generate matching error data, and the attribution relationship of the categories in the optimized clustering result is adjusted based on the matching error data. Specifically, the position of the category center is dynamically adjusted using the feature distance, and the category boundary is corrected to achieve the optimal performance of the corrected category data in terms of accuracy and consistency.

[0077] For example, in an autonomous driving scenario, by performing feature matching correction on dynamic pedestrian targets, the feature categories of different pedestrians can be accurately distinguished, ensuring the system's recognition capabilities in high-density environments.

[0078] Based on the corrected category data, the category distribution information of the target area is identified, and the category distribution information is input into the target area feature model for category optimization processing to obtain the optimized target area.

[0079] By analyzing the corrected category data, the category distribution probability map of the target area is generated, and the target area feature model is input for category optimization. The model further improves the accuracy of category division through convolution operations and iterative learning, and refines the category attribution of the boundary area to generate an optimized target area. Specifically, the parameters of the category recognition model are dynamically optimized by combining the category distribution information and feature space information, so that the optimized target area is closer to the visual target distribution of the actual scene.

[0080] For example, in precision agricultural monitoring, by optimizing the categories of field crop areas, it is possible to accurately distinguish the types of different crops and their distribution ranges, providing support for crop management and planning.

[0081] Among them, the steps of calculating the feature distance between the feature vector after dimensionality reduction and the optimized clustering result to obtain matching error data, and adaptively correcting the optimized clustering result according to the matching error data to obtain the corrected category data include: calculating the Euclidean distance, cosine similarity and Mahalanobis distance between the feature vector after dimensionality reduction and the optimized clustering result, calculating the matching error data based on the Euclidean distance and the Mahalanobis distance, and generating feature similarity data using cosine similarity.

[0082] The relationship between the reduced feature vector and the optimized clustering result is described by comprehensively using Euclidean distance, cosine similarity and Mahalanobis distance. The matching error data is generated by combining the geometric distance (Euclidean distance), directional similarity (cosine similarity) and distribution consistency (Mahalanobis distance) of the features. At the same time, feature similarity data is generated based on cosine similarity to comprehensively measure the reliability and strength of the matching.

[0083] For example, in medical images, by calculating the matching error and similarity data between the tumor area and the standard model features, a more detailed error analysis can be provided to identify potential problem areas in image diagnosis.

[0084] The dynamic weight adjustment method is applied to normalize the matching error data based on the feature similarity data to obtain normalized matching error data and feature contribution data. The feature contribution data is used to perform error compensation calculation on the normalized matching error data to obtain optimized matching error data and error distribution information.

[0085] Through the dynamic weight adjustment method, the matching error data is normalized by multiple factors according to the feature similarity data to generate standardized normalized matching error data, and the feature contribution data is used to compensate for the errors in different feature dimensions to ensure that the error proportion of high-contribution features is smaller. At the same time, the error distribution information is output for guidance of subsequent category adjustments.

[0086] For example, in space exploration missions, the category and position characteristics of unknown celestial bodies can be accurately identified by adjusting the characteristic attribution of stellar targets through error distribution information.

[0087] The error distribution information is used to adaptively adjust the optimized matching error data to obtain error correction parameters, and the category center migration adjustment is performed on the optimized clustering results according to the error correction parameters to obtain corrected category data.

[0088] By using the error distribution information to dynamically calculate the error correction parameters, the error correction parameters are combined to migrate and adjust the category center in the optimized clustering results, redefine the category boundaries, and correct the feature distribution within the category, so that the corrected category data is more representative and consistent. Specifically, the correction strength is increased in the high error area, and the original distribution is retained in the low error area to ensure the overall correction effect.

[0089] For example, in low-power edge computing devices, adjusting the category boundaries of the target area through error correction parameters can achieve efficient target recognition and classification in complex scenarios.

[0090] In step S400, the optimized target region is compared with other regions in the original image data to obtain a comparison result, and the target region in the original image data is reconstructed based on the comparison result by an image reconstruction algorithm to obtain a reconstructed target region.

[0091] By comparing the optimized target area with other areas in the original image data at the pixel level, the similarity difference between the two in terms of light intensity, texture and edge features is calculated to generate a comparison result. According to the comparison result, the image reconstruction algorithm based on sparse representation is used to accurately reconstruct the target area in the original image data; specifically, the feature point alignment technology and pixel interpolation method are combined to enhance the detail characteristics of the target area, while ensuring the natural transition between it and the surrounding background area, so as to obtain the reconstructed target area.

[0092] For example, in industrial defect detection, by comparing and analyzing the optimized defect area on the metal plate surface with the original image data, a smooth and flawless metal surface area can be effectively reconstructed for subsequent quality assessment.

[0093] The reconstructed target area is fused with other areas in the original image data to obtain fused image data, the fused image data is preprocessed to obtain processed image data, and the processed image data is compressed using an image compression algorithm to obtain compressed image data.

[0094] By fusing the reconstructed target area with other areas in the original image data, a multi-resolution fusion method is used to dynamically match the image parts to eliminate edge splicing traces. At the same time, the color balancing algorithm is used to optimize the visual consistency of the fused area to generate fused image data. Subsequently, the fused image data is preprocessed, including quantization, de-redundancy and encoding processing, to ensure the compactness and high fidelity of the processed image data in storage and transmission. Finally, the processed image data is further compressed by an image compression algorithm (such as JPEG compression or a compression algorithm based on deep learning) to obtain compressed image data.

[0095] For example, in intelligent monitoring scenarios, by fusing and compressing the moving target areas detected in dynamic scenes, high-definition but efficiently compressed video frame data can be generated to adapt to real-time monitoring needs with limited bandwidth.

[0096] In this embodiment, by collecting the original image data and noise feature data output by the sensor, the multi-scale filtering method is combined with the noise mapping relationship to perform denoising and adaptive noise reduction processing, which effectively improves the quality of the image data, and optimizes and reconstructs the artifact residue and edge area through artifact removal, edge refinement and interpolation methods, and finally obtains clear image data. At the same time, in the process of image feature extraction and regional analysis, feature stratification and domain adaptive image segmentation methods are used to combine global feature data with local feature data, extract multiple candidate target areas and regional confidence data, and select target areas and target feature data through feature fusion strategies, and use the Transformer self-attention mechanism and graph analysis method in the deep learning model to complete feature mapping, feature association and dimensionality reduction calculation, and generate a feature vector after dimensionality reduction. For the classification and optimization of the target area, a multi-classification method and a hierarchical clustering strategy are applied, and the optimized target area is finally obtained through the optimization adjustment of the category data, adaptive correction and optimization processing of the category distribution information. In the data fusion and reconstruction link, by reconstructing the target area and fusing it with other areas, high-quality image compression is achieved in combination with the image compression algorithm, which significantly improves the efficiency of data storage and transmission. The entire solution uses a multi-level, multi-step processing method, which not only significantly improves the quality and classification accuracy of image data, but also shows outstanding advantages in optimizing efficiency and reducing power consumption, providing an efficient solution for image processing of low-power MCU vision modules.

[0097] Embodiment 3: like Figure 2 As shown, the present application also provides an image processing device 10 of a low-power MCU vision module, including an acquisition module 11, a dimensionality reduction module 12, a clustering module 13 and a compression module 14.

[0098] The acquisition module 11 is mainly used to acquire the original image data collected by the sensor and perform denoising on the original image data to obtain clear image data.

[0099] The dimension reduction module 12 is mainly used to perform target recognition based on clear image data, obtain the target area, input the target area into a preset deep learning model, obtain feature data, perform dimension reduction processing on the feature data, and obtain a feature vector after dimension reduction.

[0100] The clustering module 13 is mainly used to identify several items of category data and category confidence data of the target area according to the feature vector after dimensionality reduction, cluster all the category data to generate clustering results, optimize the clustering results using the category confidence data to obtain optimized clustering results, optimize the target area according to the optimized clustering results to obtain the optimized target area.

[0101] The compression module 14 is mainly used to reconstruct the original image data using the optimized target area to obtain processed image data, and input the processed image data into the image compression algorithm for calculation to obtain compressed image data.

[0102] In this embodiment, the original image data collected by the sensor is accurately processed by the acquisition module 11, and the noise in the data is eliminated by using the multi-scale denoising technology, and the detailed characteristics of the image are retained, so as to generate clear image data, laying the foundation for subsequent processing; the dimension reduction module 12 completes the recognition of the target area and the feature extraction of the deep learning model based on the clear image data, and combines the dimension reduction algorithm to remove redundant feature data, and finally generates the feature vector after dimension reduction, which not only improves the efficiency of data processing, but also retains the key characteristics of the target; the clustering module 13 performs multi-dimensional analysis on the feature vector after dimension reduction, generates category data and performs clustering optimization, and further optimizes the characteristic performance of the target area by dynamically adjusting the matching relationship between the clustering results and the category confidence data, and obtains the accurate optimized target area; finally, the compression module 14 reconstructs the original image data in combination with the optimization characteristics of the target area, and uses an efficient image compression algorithm to compress the processed image data, reducing the burden of data storage and transmission, while retaining key image information.

[0103] Through modular design, an efficient and accurate image processing process is achieved, which significantly improves the image processing capability and power consumption efficiency of the low-power MCU vision module in complex scenarios, and provides reliable technical support for visual tasks in edge computing.

[0104] It should be noted that technicians in the relevant technical field can clearly understand that for the convenience and simplicity of description, the specific working process of the above-described device and each module can refer to the corresponding process in the aforementioned embodiment 1, and will not be repeated here.

[0105] The structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantive technical significance. Any structural modification, change in proportion or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical contents disclosed by the present invention.

[0106] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An image processing method for a low-power MCU visual module, characterized in that: include: Acquire the original image data collected by the sensor, and perform denoising on the original image data to obtain clear image data; Performing target recognition based on the clear image data to obtain a target area, inputting the target area into a preset deep learning model to obtain feature data, and performing dimensionality reduction processing on the feature data to obtain a feature vector after dimensionality reduction; Identifying several items of category data and category confidence data of the target area according to the feature vector after dimension reduction, performing clustering processing on all the category data to generate a clustering result, optimizing the clustering result using the category confidence data to obtain an optimized clustering result, and optimizing the target area according to the optimized clustering result to obtain an optimized target area; The original image data is reconstructed using the optimized target area to obtain processed image data, and the processed image data is input into an image compression algorithm for calculation to obtain compressed image data.

2. The image processing method of the low-power MCU visual module according to claim 1, characterized in that: The step of acquiring the original image data collected by the sensor and performing denoising on the original image data to obtain clear image data includes: Collecting raw image data output by the sensor and noise characteristic data corresponding to the raw image data, and denoising the raw image data and the noise characteristic data using a multi-scale filtering method to obtain image data after preliminary denoising and residual noise data; Constructing a noise mapping relationship through the residual noise data, and performing adaptive noise reduction processing on the image data after preliminary denoising based on the noise mapping relationship to obtain target denoised image data; Performing artifact removal on the target denoised image data to obtain optimized image data and artifact feature data, calculating gradient change information of the optimized image data, performing edge refinement processing on the gradient change information and the artifact feature data to obtain clear image data.

3. The image processing method of the low-power MCU visual module according to claim 2, characterized in that: The step of removing artifacts from the target denoised image data to obtain optimized image data and artifact feature data, calculating gradient change information of the optimized image data, and performing edge thinning processing on the gradient change information and the artifact feature data to obtain clear image data includes: Performing spectral decomposition on the denoised image data of the target based on a frequency domain analysis method to obtain component data, extracting artifact feature data based on the component data, and performing artifact suppression on the component data based on the artifact feature data using a directional filtering method to obtain artifact suppressed image data and artifact residual data; Calculating local gradient change information of the artifact-suppressed image data, and adaptively compensating the artifact residual data based on the local gradient change information to obtain optimized image data; An interpolation method is applied to reconstruct the edges of the optimized image data to obtain edge-optimized image data and edge error information, the overall texture information of the edge-optimized image data is calculated, and texture enhancement processing is performed on the edge-optimized image data based on the overall texture information and edge error information to obtain clear image data.

4. The image processing method of the low-power MCU visual module according to claim 1, characterized in that: The steps of performing target recognition based on the clear image data to obtain a target area, inputting the target area into a preset deep learning model to obtain feature data, and performing dimensionality reduction processing on the feature data to obtain a feature vector after dimensionality reduction include: Performing feature stratification on the clear image data to obtain global feature data and local feature data, applying a domain adaptive image segmentation method to perform region extraction on the global feature data and the local feature data to obtain a plurality of candidate target regions and region confidence data; Based on the regional confidence data, a plurality of candidate target regions are screened using a preset feature fusion strategy to obtain target regions and target feature data; Inputting the target area and the target feature data into a preset deep learning model, applying a feature mapping method of a Transformer self-attention mechanism in the deep learning model to perform feature mapping calculation on the target area and the target feature data to obtain an original feature vector and spatial feature relationship data; Based on the spatial feature relationship data, a graph analysis method is used to perform feature association on the original feature vector to obtain an associated feature vector and feature importance data, and the associated feature vector and feature importance data are used to perform dimensionality reduction calculation on the original feature vector to obtain a feature vector after dimensionality reduction.

5. The image processing method of the low-power MCU visual module according to claim 1, characterized in that: The steps of identifying several items of category data and category confidence data of the target area according to the feature vector after dimension reduction, clustering all the category data to generate clustering results, optimizing the clustering results using the category confidence data to obtain optimized clustering results, and optimizing the target area according to the optimized clustering results to obtain the optimized target area include: Inputting the reduced-dimensional feature vector into a preset target region feature model, applying a multi-classification method in the target region feature model to classify the target region based on the reduced-dimensional feature vector, and obtaining a plurality of category data and category confidence data; Performing hierarchical clustering on all the category data to obtain clustering results and category center data, and optimizing and adjusting the clustering results using the category confidence data and the category center data to obtain optimized clustering results; Calculating the characteristic distance between the feature vector after dimension reduction and the optimized clustering result to obtain matching error data, and adaptively correcting the optimized clustering result according to the matching error data to obtain corrected category data; Based on the corrected category data, category distribution information of the target area is identified, and the category distribution information is input into the target area feature model for category optimization processing to obtain an optimized target area.

6. The image processing method of the low-power MCU visual module according to claim 5, characterized in that: The step of calculating the characteristic distance between the feature vector after dimension reduction and the optimized clustering result to obtain matching error data, and adaptively correcting the optimized clustering result according to the matching error data to obtain corrected category data includes: Calculating the Euclidean distance, cosine similarity and Mahalanobis distance between the feature vector after dimension reduction and the optimized clustering result, calculating matching error data based on the Euclidean distance and the Mahalanobis distance, and generating feature similarity data using the cosine similarity; Applying a dynamic weight adjustment method to normalize the matching error data based on the feature similarity data to obtain normalized matching error data and feature contribution data, and using the feature contribution data to perform error compensation calculation on the normalized matching error data to obtain optimized matching error data and error distribution information; The error distribution information is used to adaptively adjust the optimized matching error data to obtain error correction parameters, and the optimized clustering results are adjusted for category center migration according to the error correction parameters to obtain corrected category data.

7. The image processing method of the low-power MCU visual module according to claim 1, characterized in that: The step of reconstructing the original image data using the optimized target area to obtain processed image data, and inputting the processed image data into an image compression algorithm for calculation to obtain compressed image data comprises: Comparing the optimized target region with other regions in the original image data to obtain a comparison result, and reconstructing the target region in the original image data based on the comparison result by an image reconstruction algorithm to obtain a reconstructed target region; The reconstructed target area is fused with other areas in the original image data to obtain fused image data, the fused image data is preprocessed to obtain processed image data, and the processed image data is compressed using an image compression algorithm to obtain compressed image data.

8. An image processing device of a low-power MCU visual module, characterized in that: include: An acquisition module is used to acquire the original image data collected by the sensor and perform denoising on the original image data to obtain clear image data; A dimension reduction module is used to perform target recognition based on the clear image data to obtain a target area, input the target area into a preset deep learning model to obtain feature data, perform dimension reduction processing on the feature data, and obtain a feature vector after dimension reduction; A clustering module, used to identify several items of category data and category confidence data of the target area according to the feature vector after dimension reduction, perform clustering processing on all the category data to generate a clustering result, optimize the clustering result using the category confidence data to obtain an optimized clustering result, and optimize the target area according to the optimized clustering result to obtain an optimized target area; The compression module is used to reconstruct the original image data using the optimized target area to obtain processed image data, and input the processed image data into an image compression algorithm for calculation to obtain compressed image data.

Images