Big data image processing method and system
By employing big data image processing methods, combined with image preprocessing, complexity analysis, and parallel processing of lightweight models, the problems of low image processing efficiency and long training time in existing technologies are solved, achieving efficient and accurate image quality assessment and target detection.
Patent Information
- Application Number
- CN202510908928.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-02
- Publication Date
- 2025-11-07
AI Technical Summary
Existing image processing methods are labor-intensive and time-consuming when processing ultra-large-scale images. Traditional algorithms are inefficient and struggle to effectively address image quality degradation caused by factors such as ambient lighting and noise interference.
We employ big data image processing methods, including image preprocessing, complexity analysis, lightweight model parallel processing, and cross-node feature fusion. We utilize image enhancement, segmentation, reconstruction, and denoising techniques, combined with lightweight models and multi-scale peak signal-to-noise ratio detection, to output high-quality images through parallel processing and feature fusion.
It significantly reduces computing resource consumption, improves image processing efficiency, enhances the accuracy and quality of target detection, reduces energy consumption, and achieves higher performance and lower power consumption.
Smart Images
Figure CN120912898A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of image processing, and particularly relates to a big data image processing method and system. BACKGROUND
[0002] In daily life and production activities, images are very important information transmission media. However, in the process of collecting and transmitting images, there are adverse factors affecting imaging, such as environmental light, noise interference, hardware support, and excessive compression, which result in poor data quality and make the useful information in the image unable to be normally read and recognized. Therefore, researchers have developed various image processing methods, that is, processing images before transmitting them to ensure their clarity and accuracy.
[0003] However, the existing methods have the following problems: in the process of processing images, confirming whether all feature parameters are qualified greatly increases the workload and reduces the work efficiency; and the traditional image processing algorithm (such as CNN) takes too long to train when processing super large scale images due to too large data volume. SUMMARY
[0004] In view of the deficiencies of the prior art, the present application aims to provide a big data image processing method and system to effectively reduce the workload and improve the image processing efficiency.
[0005] To achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows: In a first aspect, the present application provides a big data image processing method, the key of which lies in comprising the following steps: S1, acquiring an image to be processed and performing image preprocessing; S2, dynamically dividing the image into a plurality of image sub-regions based on the image complexity analysis result of the preprocessed image; S3, using a pre-trained lightweight model to process the image sub-regions in parallel to extract the edge information of the target, performing edge recognition on the target in the image sub-region according to the edge information and the pixel point features of the image sub-regions, and then obtaining the image pixel line of the target edge contour; S4, determining the edge fluctuation cost of the target in the image sub-region through the image pixel line and the shape feature and spatial relationship feature of the target, performing quality recognition on the pixel points in the image sub-region according to the edge fluctuation cost, and obtaining the edge detection confidence information of the target; S5, extracting the target in each image sub-region according to the edge detection confidence information of the target; S6, integrating the processing results of each image sub-region using a cross-node feature fusion model, and outputting a complete global image.
[0006] Further, the image preprocessing in step S1 includes: Image enhancement: enhancing the to-be-processed image by using image enhancement technology to obtain an enhanced image; Image segmentation: processing the enhanced image by using instance segmentation technology to obtain a plurality of segmentation regions, and identifying each segmentation region as a patch image; Image reconstruction: performing super-resolution reconstruction on any one patch image by using super-resolution reconstruction technology to obtain a reconstructed image; Image denoising: performing multi-scale peak signal-to-noise ratio detection on the reconstructed image, and optimizing the corresponding reconstructed image according to the peak signal-to-noise ratio detection result to obtain a denoised image.
[0007] Further, the multi-scale peak signal-to-noise ratio detection includes the following steps: performing multi-scale division on the reconstructed image to obtain a plurality of scale reconstructed image blocks; respectively performing peak signal-to-noise ratio detection on the reconstructed image blocks under multiple scales to obtain and determine the final peak signal-to-noise ratio detection result according to the detection results under multiple scales.
[0008] Further, the optimization of the corresponding reconstructed image according to the peak signal-to-noise ratio detection result to obtain a denoised image includes the following steps: determining whether the peak signal-to-noise ratio detection result is greater than a first preset threshold, if greater, directly identifying the reconstructed image block as a high-quality image; if less than and greater than a second preset threshold, using any one of the median filter denoising methods to optimize the corresponding reconstructed image block; if the peak signal-to-noise ratio detection result is less than the second threshold, using a denoising method based on a convolutional neural network to optimize the corresponding reconstructed image block.
[0009] Further, the image complexity analysis in step S2 includes the following steps: calculating the average gradient value and the average brightness value of each pixel in the image; calculating the edge point average gradient value and the edge point average brightness value of each edge pixel in the image; calculating the image complexity analysis result according to the average gradient value, the average brightness value, the edge point average gradient value and the edge point average brightness value.
[0010] Further, the pre-trained lightweight model in step S3 is a pruned MobileNetV3, and the pruning rate is adaptively determined according to the peak signal-to-noise ratio of the image sub-region.
[0011] Further, the image pixel line of the target edge contour obtained in step S3 includes the following steps: extracting image contour features of each of the image sub-regions; determining image edge features of the target according to all the image contour features; determining image edge information of the target through the image edge features and the image sub-regions; obtaining pixel point features of the image sub-regions; determining edge recognition information of the target according to the image edge information and the pixel point features of the image sub-regions; determining a plurality of image contour lines according to the edge recognition information and the image sub-regions; determining image pixel lines of the target edge contour through all the image contour lines.
[0012] Further, the step S4 of obtaining the detection confidence information of the target comprises the following steps: obtaining shape features and spatial relationship features of the target; calculating fluctuation entropy of the shape features and the spatial relationship features; determining edge fluctuation cost of the target in the image sub-region according to the fluctuation entropy and the image pixel lines; calculating pixel value data of each pixel point in the image sub-region; determining edge detection confidence index of each pixel point through the pixel value data and the image pixel lines; obtaining edge detection confidence information of the target through the edge fluctuation cost and the edge detection confidence information of the pixel point.
[0013] Further, the cross-node feature fusion model is deployed on an FPGA, and a pipeline ping-pong operation is used to temporarily store image data.
[0014] In a second aspect, the present application provides a big data image processing system for implementing the method of the first aspect, comprising: an image preprocessing module for obtaining an image to be processed and performing image preprocessing; an image division module for dynamically dividing the image into a plurality of image sub-regions according to the image complexity analysis result of the preprocessed image; an edge recognition module for parallel processing of the image sub-regions using a pre-trained lightweight model to extract edge information of the target, performing edge recognition on the target in the image sub-region according to the edge information and pixel point features of the image sub-regions, and further obtaining image pixel lines of the target edge contour; The data processing module is used to determine the edge fluctuation cost of the target in the sub-region of the image by using the image pixel lines and the shape and spatial relationship features of the target, and to perform quality recognition on the pixels in the sub-region of the image based on the edge fluctuation cost to obtain the edge detection confidence information of the target. The target extraction module is used to extract targets in each of the image sub-regions based on the edge detection confidence information of the targets; The feature fusion module is used to integrate the processing results of each image sub-region using a cross-node feature fusion model to output a complete global image.
[0015] The significant effects of this invention are: 1. This invention first preprocesses the image to acquire it using image enhancement-based instance segmentation technology, which can more accurately obtain images containing the target. Simultaneously, super-resolution reconstruction technology is used to reconstruct the image, resulting in a higher-quality target image. Based on this, this invention uses a multi-scale peak signal-to-noise ratio (PSNR) detection method to detect the image, achieving accurate image quality assessment. For low-quality images, the product area image is optimized based on the detection results, significantly reducing computational resource consumption while ensuring optimization quality. Then, based on the image complexity analysis results, the image is divided into blocks to ignore redundant information and improve processing efficiency. Afterwards, targets in each image sub-region are extracted through parallel processing and integrated to output a complete global image, greatly reducing workload and improving efficiency compared to existing technologies.
[0016] 2. In the target extraction process, the fluctuation degree of the edge recognition line of the target shape features and spatial relationship features in the image is first determined, and then the edge fluctuation cost is obtained to realize the edge of the target in the image, thereby suppressing the influence of edge fluctuation on the target detection. Secondly, the quality recognition of the pixels in the image sub-region is performed by the edge fluctuation cost to obtain the target edge detection confidence information, which realizes the accurate positioning of the target in the image and finds out the small defects in the target image, thereby improving the target detection quality.
[0017] 3. The lightweight model adopted in this invention reduces the number of model parameters by using pruning and quantization techniques, overcoming the defect of excessive training time caused by excessive data volume, effectively improving running efficiency, ensuring continuous optimization and stability of the model, and enabling real-time inference services to be deployed in the cloud or on edge devices.
[0018] 4. In the COCO dataset test, the present invention achieved an mAP of 78.9% while reducing energy consumption by 22%, demonstrating higher performance and lower power consumption compared to existing technologies. Attached Figure Description
[0019] Figure 1 is a flowchart of the method of the present application; Figure 2 is a flowchart of the image preprocessing of the present application; Figure 3 is a flowchart of the image complexity analysis of the present application; Figure 4 is a flowchart of the acquisition of the image pixel line of the present application; Figure 5 is a flowchart of the acquisition of the target detection confidence information of the present application; Figure 6 is a structural schematic diagram of the system of the present application. DETAILED DESCRIPTION
[0020] The specific embodiments of the present application and the working principles will be further described in detail below with reference to the accompanying drawings. EMBODIMENT
[0021] As shown in Figure 1 , the present embodiment proposes a big data image processing method, and the specific steps are as follows: S1, acquiring a to-be-processed image and performing image preprocessing; When specifically implemented, the flowchart of the image preprocessing of the present application is as shown in Figure 2 , and specifically includes the following steps: Step S11, image enhancement: using image enhancement technology to perform enhancement processing on the to-be-processed image to obtain an enhanced image; Step S12, image segmentation: using instance segmentation technology to process the enhanced image to obtain a plurality of segmentation regions, and identifying each segmentation region as a block image; Step S13, image reconstruction: using super-resolution reconstruction technology to perform super-resolution reconstruction on any one block image to obtain a reconstructed image; Step S14, image denoising: performing multi-scale peak signal-to-noise ratio detection on the reconstructed image, and performing optimization on the corresponding reconstructed image according to the peak signal-to-noise ratio detection result to obtain a denoised image.
[0022] In the above preprocessing process, the image is obtained by using the instance segmentation technology based on image enhancement, so that the image containing the target is more accurately obtained; meanwhile, the image is reconstructed by using the super-resolution reconstruction technology, so that a higher quality target image is obtained; based on this, the image is detected by using the multi-scale peak signal-to-noise ratio detection method, so that the accurate evaluation of the image quality is realized; for the non-high-quality image, the image of the commodity area is optimized according to the detection result, so that the calculation resource consumption is significantly reduced on the premise of ensuring the optimization quality; then, the image is divided into blocks based on the image complexity analysis result, so as to ignore the redundant information and improve the processing efficiency.
[0023] In some embodiments, the multi-scale peak signal-to-noise ratio detection comprises the following steps: The reconstructed image is divided into multiple scales to obtain reconstructed image blocks of multiple scales; The peak signal-to-noise ratio of the reconstructed image blocks under multiple scales is detected respectively to obtain the final peak signal-to-noise ratio detection result according to the detection results under multiple scales.
[0024] Further, the optimization of the corresponding reconstructed image to obtain the denoising image according to the peak signal-to-noise ratio detection result can be realized by the following steps: If the peak signal-to-noise ratio detection result is greater than a first preset threshold, the reconstructed image block is directly identified as a high-quality image; If it is less than and greater than a second preset threshold, the corresponding reconstructed image block is optimized by using any one of the median filtering denoising methods; If the peak signal-to-noise ratio detection result is less than the second threshold, the corresponding reconstructed image block is optimized by using the denoising method based on the convolutional neural network.
[0025] It can be understood that, based on the above peak signal-to-noise ratio detection and image optimization, if the peak signal-to-noise ratios of the reconstructed image blocks under all scales are high, the reconstructed image block is directly identified as a high-quality image without optimization; if the peak signal-to-noise ratios of the reconstructed image blocks under most scales are high, the reconstructed image block is denoised and optimized by using the ordinary denoising method; if the peak signal-to-noise ratios of the reconstructed image blocks under a small part of scales are high, the reconstructed image block is denoised and optimized by using the denoising method of deep learning.
[0026] S2, based on the image complexity analysis result of the preprocessed image, the image is dynamically divided into a plurality of image sub-regions; In the specific implementation process, the flow of the image complexity analysis can refer to the attached Figure 3 , comprising the following steps: S2.1, convert the pre-processed image into a gray image, calculate the gradient value and brightness value of each pixel by using a sliding window, and then calculate the average gradient value and average brightness value of each pixel in the image according to the gradient value and brightness value of each pixel; S2.2, generate an edge picture according to the gradient value of the pixel and a preset threshold, calculate the gradient value and brightness value of each edge pixel point according to the number of edge pixel points of the edge picture, and then calculate the edge point average gradient value and edge point average brightness value of each edge pixel in the image; S2.3, according to experience, the four indicators obtained by calculation are weighted, the value of the weight is set according to the value range of TP, and the value of EP is controlled according to the value of TP and Param; then the image complexity analysis result is calculated according to the weighted average gradient value, average brightness value, edge point average gradient value and edge point average brightness value.
[0027] The embodiment divides the image based on the image complexity analysis result, effectively ignores the redundant information, and improves the processing efficiency.
[0028] S3, a pre-trained lightweight model is used to process the image sub-regions in parallel to extract the edge information of the target, and the edge information and the pixel point features of the image sub-regions are used to identify the target in the image sub-regions, and then the image pixel line of the target edge contour is obtained. In this example, the pre-trained lightweight model is a pruned MobileNetV3, and the pruning rate is adaptively determined according to the peak signal-to-noise ratio of the image sub-region.
[0029] The MobileNetV3 neural network introduces a new nonlinear h-swish, which is faster in calculation and more friendly to quantization. It is a 28-layer network of depth convolutional neural network with depth separable convolution kernel as the basic structure. A depth separable kernel is composed of a depth convolution kernel and a point convolution kernel. The MobileNetV3 neural network includes 1 convolution layer, 13 depth separable layers, a global average pooling layer and a fully connected output layer, and the network has no pooling layer; the number of convolution layers is selected, and the convolution kernel of the layer is replaced by a dilated convolution kernel, the quality of the learned features is improved by increasing the receptive field of the convolution kernel, and the classification accuracy is further improved; wherein, the replaced convolution kernel is one or two layers of high-resolution input features with high resolution input features. Before training, representative training samples are extracted from the fused data and proportionally divided into a training set, a validation set and a test set. Hyperparameters are set, model training is performed, data augmentation and regularization techniques are applied during training to improve the generalization ability of the model, prevent overfitting, and evaluate the model performance using the test set. The best model is selected for deployment based on accuracy, recall and F1 score key indicators. This process ensures that the model has high reliability and effectiveness in various scenarios and conditions, providing stable support for the practical application of the system. In this example, the MobileNetV3 model training steps are as follows: input the training set image into the MobileNetv3 neural network model; train an initial model on each training set in the MobileNetv3 neural network model, and test the initial model on the corresponding test set, calculate and save the segmentation evaluation standard precision of the initial model.
[0030] The lightweight model adopted in the embodiment of the application reduces the number of model parameters by using pruning and quantization techniques, overcomes the defect of excessive training time caused by excessive data volume, effectively improves the running efficiency, and ensures the continuous optimization and stability of the model. It can be deployed on the cloud or edge device to provide real-time inference services. At the same time, by parallel processing, the target in each image sub-region is extracted and a complete global image is output after integration, which greatly reduces the workload and improves the work efficiency compared with the prior art.
[0031] As shown in Figure 4 The image pixel line of the target edge contour obtained by the embodiment includes the following steps: Step S3.1, extract the image contour features of each image sub-region; When implemented, the extracted image contour features can be implemented in the following manner: select an image sub-region as a selected image, extract the contour size of the target in the selected image sub-region by using the contour detection algorithm in the prior art, and use the contour size as the image contour feature of the selected image sub-region. Continue to determine the image contour features of the remaining image sub-regions.
[0032] Step S3.2, determine the image edge features of the target according to all the image contour features; In this example, the image edge features of the target are determined according to all the image contour features, which can be implemented in the following manner: the average size of the contour size in all image edge features is used as the image edge size of the target, and the image edge size is used as the image edge feature, wherein the image edge feature represents the edge information of the target Step S3.3, determine the image edge information of the target through the image edge features and the image sub-regions; In this example, the image edge information of the target is determined by the image edge features and the image sub-regions, which can be achieved by using a shape matching algorithm in the prior art in combination with the image edge features and the image sub-regions. In this embodiment, the shape matching algorithm is the Hu moment algorithm, which matches the shape of the target screw by calculating seven invariant moments. In other embodiments, the image edge information of the target can be determined in other ways, which are not limited here.
[0033] It should be noted that the image edge information in this application is information reflecting the contour, shape, boundary and other features of the target. The image edge information is used to detect the regions where the pixel values change significantly in the image, which usually correspond to the boundaries of the target and the texture features outside.
[0034] Step S3.4, obtaining the pixel point features of the image sub-regions; In specific implementation, the pixel point features in the image sub-regions can be obtained by using the OpenCV library of Python in the prior art to read the total number of pixel points in each image sub-region, and taking all the total number of pixel points as the pixel point features of the image sub-regions. The pixel point features represent the total number of pixel points in the image sub-regions.
[0035] Step S3.5, determining the edge recognition information of the target according to the image edge information and the pixel point features of the image sub-regions; In this application, the edge recognition of the target in the image sub-regions is performed according to the image edge information and the pixel point features of the image sub-regions, that is, the edge recognition information of the target is determined according to the image edge information and the pixel point features of the image sub-regions, and a plurality of image contour lines are determined according to the edge recognition information and the image sub-regions. The edge recognition represents the segmentation of the edge of the target in the image sub-regions.
[0036] Step S3.6, determining a plurality of image contour lines according to the edge recognition information and the image sub-regions; Step S3.7, determining the image pixel line of the edge contour of the target by all the image contour lines. The image pixel line is a reliable pixel line reflecting the edge contour of the target, which is used to segment the target in the image sub-regions.
[0037] S4, determining the edge fluctuation cost of the target in the image sub-regions by the image pixel line and the shape features and spatial relationship features of the target, performing quality recognition on the pixel points in the image sub-regions according to the edge fluctuation cost, and obtaining the edge detection confidence information of the target; Referring to FIG. 4, the edge detection confidence information of the target is obtained by the image pixel line and the shape features and spatial relationship features of the target. Figure 5In the embodiment of the present application, the obtaining of the detection confidence information of the target comprises the following steps: S4.1, obtaining the shape feature and the spatial relationship feature of the target by using an existing algorithm; S4.2, calculating the fluctuation entropy of the shape feature and the spatial relationship feature; S4.3, determining the edge fluctuation cost of the target in the image sub-region according to the fluctuation entropy and the image pixel line, specifically: multiplying the image pixel line by the fluctuation entropy, taking the region between the multiplied line and the image pixel line as the confidence fluctuation region, and taking the confidence fluctuation region as the confidence fluctuation region, then taking the value obtained by dividing the image pixel line by the fluctuation entropy as the logarithm operation with base 2, multiplying the value obtained by the logarithm operation by the confidence fluctuation region in the confidence fluctuation information, and taking the multiplied region as the edge fluctuation cost of the target, and in other embodiments, other ways can also be used for determination, which is not limited here; S4.4, calculating the pixel value data of each pixel point in the image sub-region, specifically: obtaining the pixel value set of each sub-region image in the image sub-region set by using the OpenCV library of the Python software in the prior art, wherein the pixel value set represents the set of all pixel values, and taking the set of all pixel value sets as the pixel value data in the image sub-region set; S4.5, determining the edge detection confidence index of each pixel point by using the pixel value data and the image pixel line; Selecting one image in the image sub-region set, extracting the region surrounded by the image pixel line in the image, extracting all pixel values of the surrounded region from the pixel value data, uniformly dividing the surrounded region into 9 sub-regions, calculating the standard deviation, mean value and histogram of the pixel values in each sub-region according to the extracted all pixel values, and taking the set of the standard deviation, mean value and histogram as the edge detection confidence information set of each pixel point in the image, wherein the edge detection confidence information represents the confidence information of the pixel point in the target detection process, and is used for judging whether the pixel point is the target in the image sub-region set.
[0038] S4.6, obtaining the edge detection confidence information of the target by using the edge fluctuation cost and the edge detection confidence information of the pixel point.
[0039] The target edge detection confidence information obtained by the edge fluctuation cost and the edge detection confidence information of the pixel point can be obtained by a machine learning model, for example: target edge detection confidence information = edge fluctuation cost * A + edge detection confidence information of the pixel point * B, wherein A and B are weight coefficients, A and B can be determined after adaptive adjustment according to experience or a learning process, and in other embodiments, other ways of determination can also be used, which are not limited here.
[0040] In the target extraction process, the edge fluctuation cost is obtained by first determining the fluctuation degree of the edge recognition line of the target shape feature and the spatial relationship feature in the image, and then the edge of the target in the image is determined, so as to suppress the influence of the edge fluctuation on the detection target. Secondly, the edge fluctuation cost is used to identify the quality of the pixel points in the image sub-region, and the target edge detection confidence information is obtained, so as to realize accurate positioning of the target in the image, find out the small defects and other situations in the target image, and improve the detection quality of the target.
[0041] S5, extracting the target in each image sub-region according to the edge detection confidence information of the target; S6, integrating the processing results of each image sub-region by using a cross-node feature fusion model, and outputting a complete global image.
[0042] In the embodiment, the image preprocessing model and the cross-node feature fusion model are deployed in the FPGA, and the pipeline ping-pong operation is used for temporary storage of image data. The FPGA is divided into two FIFO regions, FIFO0 and FIFO1, which are used as temporary storage regions of image data in a ping-pong operation mode. The data writing and reading sequence is in a first-in first-out mode. Preferably, the pipeline ping-pong operation means that initially, FIFO0 starts to store data, and FIFO1 is empty at this time. When FIFO0 is full, FIFO1 continues to store data, and FIFO0 performs reading operation at this time. Then, when FIFO1 is full, FIFO0 performs data storage again, and FIFO1 performs reading operation. The above operations are repeated.
[0043] In this embodiment, the image complexity analysis model and the image division model in step S2 are deployed on the CPU, and the target detection process in steps S3-S5 is deployed on the GPU. Thus, the CPU host unit as the host is responsible for processing resource management, task allocation and the like, and the GPU and the FPGA are responsible for processing computationally intensive tasks, so as to give full play to the advantages of the CPU in management and control, the advantages of the GPU in parallel processing, the performance-power ratio of the FPGA and the advantage of flexible configuration, and to adapt to different application scenarios, meet different types of task requirements, and effectively solve the problem of low efficiency of the CPU and the problem of very large power consumption of the GPU which cannot realize arbitrary precision operation and cause spatial operation waste.
[0044] Through testing the method described in the embodiment on the COCO data set, the mAP reaches 78.9%, and the energy consumption is reduced by 22%. It can be seen that the present application has higher performance and lower power consumption than the prior art, and achieves the purpose of the application. Embodiment
[0045] Referring to the accompanying Figure 6 , the present embodiment proposes a big data image processing system for implementing the method described in embodiment 1, comprising: An image preprocessing module for acquiring an image to be processed and performing image preprocessing; An image division module for dynamically dividing the image into a plurality of image sub-regions according to the image complexity analysis result of the preprocessed image; An edge recognition module for using a pre-trained lightweight model to perform parallel processing on the image sub-regions to extract edge information of the target, and performing edge recognition on the target in the image sub-region according to the edge information and the pixel point features of the image sub-regions, and then obtaining an image pixel line of the target edge contour; A data processing module for determining the edge fluctuation cost of the target in the image sub-region through the image pixel line and the shape feature and the spatial relationship feature of the target, and performing quality recognition on the pixel points in the image sub-region according to the edge fluctuation cost to obtain edge detection confidence information of the target; A target extraction module for extracting the target in each image sub-region according to the edge detection confidence information of the target; A feature fusion module for integrating the processing results of each image sub-region using a cross-node feature fusion model to output a complete global image.
[0046] To sum up, the present application firstly pre-processes the image to obtain the image containing the target more accurately by using the instance segmentation technology based on image enhancement; meanwhile, the image is reconstructed by using the super-resolution reconstruction technology to obtain the target image with higher quality; based on this, the present application detects the image by using the multi-scale peak signal-to-noise ratio detection method to realize the accurate evaluation of the image quality; for the non-high-quality image, the commodity region image is optimized according to the detection result, thereby significantly reducing the consumption of computing resources under the premise of ensuring the optimization quality; then the image is divided into blocks based on the image complexity analysis result to ignore the redundant information and improve the processing efficiency; then the target in each image sub-region is extracted by parallel processing and a complete global image is output after integration, which greatly reduces the workload and improves the work efficiency compared with the prior art. Meanwhile, the lightweight model reduces the model parameter amount by using the pruning and quantization technology, overcomes the defect that the training time is too long due to the excessive data amount, effectively improves the operation efficiency and ensures the continuous optimization and stability of the model.
[0047] The technical solutions provided by the present application are described in detail above. The principles and implementation modes of the present application are described by using specific examples in this paper, and the above examples are only used to help understand the method of the present application and its core idea. It should be pointed out that, for ordinary skilled persons in the technical field, some improvements and modifications can be made to the present application without departing from the principles of the present application, and these improvements and modifications also fall within the protection scope of the claims of the present application.
Claims
1. A big data image processing method, characterized in that, The method comprises the following steps: S1, obtaining an image to be processed and performing image preprocessing; S2, based on the image complexity analysis result of the preprocessed image, the image is dynamically divided into a plurality of image sub-regions; S3, a pre-trained lightweight model is used to process the image sub-regions in parallel to extract the edge information of the target, and the edge information and the pixel point features of the image sub-regions are used to identify the edge of the target in the image sub-region, and then the image pixel line of the target edge contour is obtained; S4, the edge fluctuation cost of the target in the image sub-region is determined by the image pixel line and the shape feature and the spatial relationship feature of the target, and the quality of the pixel points in the image sub-region is identified according to the edge fluctuation cost, and the edge detection confidence information of the target is obtained; S5, according to the edge detection confidence information of the target, the target in each image sub-region is extracted; S6, the processing results of each image sub-region are integrated by using a cross-node feature fusion model, and a complete global image is output.
2. The big data image processing method of claim 1, wherein: The image preprocessing in step S1 comprises: Image enhancement: using image enhancement technology to enhance the image to be processed to obtain an enhanced image; Image segmentation: using instance segmentation technology to process the enhanced image to obtain a plurality of segmentation regions, and identifying each segmentation region as a block image; Image reconstruction: using super-resolution reconstruction technology to perform super-resolution reconstruction on any block image to obtain a reconstructed image; Image denoising: multi-scale peak signal-to-noise ratio detection is performed on the reconstructed image, and the corresponding reconstructed image is optimized according to the peak signal-to-noise ratio detection result to obtain a denoised image.
3. The big data image processing method of claim 2, wherein: The multi-scale peak signal-to-noise ratio detection comprises the following steps: The reconstructed image is divided into multiple scales to obtain reconstructed image blocks of multiple scales; The peak signal-to-noise ratio of the reconstructed image blocks under multiple scales is detected respectively to obtain the final peak signal-to-noise ratio detection result according to the detection results under multiple scales.
4. The big data image processing method of claim 2 or 3, characterized in that: The optimization of the corresponding reconstructed image according to the peak signal-to-noise ratio detection result to obtain the denoised image comprises the following steps: If the peak signal-to-noise ratio detection result is greater than a first preset threshold, the reconstructed image block is directly identified as a high-quality image; If it is less than and greater than a second preset threshold, any one of the median filter denoising methods is used to optimize the corresponding reconstructed image block; If the peak signal-to-noise ratio detection result is less than the second threshold, the corresponding reconstructed image block is optimized by using the denoising method based on the convolutional neural network.
5. The big data image processing method of claim 1, wherein: The image complexity analysis in step S2 comprises the following steps: Calculate the average gradient value and the average brightness value of each pixel in the image; Calculate the average gradient value and the average brightness value of each edge pixel in the image; The image complexity analysis result is calculated according to the average gradient value, the average brightness value, the average gradient value of the edge point and the average brightness value of the edge point.
6. The big data image processing method of claim 1, wherein: The pre-trained lightweight model in step S3 is a pruned MobileNetV3, and the pruning rate is adaptively determined according to the peak signal-to-noise ratio of the image sub-region.
7. The big data image processing method of claim 1, wherein: The image pixel line of the target edge contour obtained in step S3 comprises the following steps: extracting image contour features of each of the image sub-regions; determining image edge features of the target according to all the image contour features; determining image edge information of the target through the image edge features and the image sub-regions; obtaining pixel point features of the image sub-regions; determining edge recognition information of the target according to the image edge information and the pixel point features of the image sub-regions; determining a plurality of image contour lines according to the edge recognition information and the image sub-regions; determining image pixel lines of the target edge contour through all the image contour lines.
8. The big data image processing method of claim 1, wherein: The detection confidence information of the target obtained in step S4 includes the following steps: obtaining shape features and spatial relationship features of the target; calculating fluctuation entropy of the shape features and the spatial relationship features; determining edge fluctuation cost of the target in the image sub-region according to the fluctuation entropy and the image pixel lines; calculating pixel value data of each pixel point in the image sub-region; determining edge detection confidence index of each pixel point through the pixel value data and the image pixel lines; obtaining edge detection confidence information of the target through the edge fluctuation cost and the edge detection confidence information of the pixel point.
9. The big data image processing method of claim 1, wherein: The cross-node feature fusion model is deployed on an FPGA, and a pipeline ping-pong operation is used for temporary storage of image data.
10. A big data image processing system for implementing the method of any one of claims 1-9, characterized by, It includes: an image preprocessing module for obtaining an image to be processed and performing image preprocessing; an image division module for dynamically dividing the image into a plurality of image sub-regions according to image complexity analysis results of the preprocessed image; an edge recognition module for parallel processing of the image sub-regions using a pre-trained lightweight model to extract edge information of the target, performing edge recognition on the target in the image sub-region according to the edge information and pixel point features of the image sub-regions, and obtaining image pixel lines of the target edge contour; a data processing module for determining edge fluctuation cost of the target in the image sub-region through the image pixel lines and shape features and spatial relationship features of the target, performing quality recognition on pixel points in the image sub-region according to the edge fluctuation cost, and obtaining edge detection confidence information of the target; a target extraction module for extracting the target in each of the image sub-regions according to the edge detection confidence information of the target; a feature fusion module for integrating processing results of each of the image sub-regions using a cross-node feature fusion model, and outputting a complete global image.