Image deep learning preprocessing method based on feature compression

Abstract

The application discloses a picture deep learning preprocessing method based on feature compression, and relates to the technical field of computer vision and image processing.The preprocessing method comprises the following steps: frequency domain decomposition, parameter dynamic adjustment compression encoding and reconstruction of an approximate original picture.The application has the advantages that: through frequency domain decomposition and selective compression, while the core visual information (such as object contour and texture distribution) of the picture is retained, the calculation burden of the deep learning model is greatly reduced; compared with traditional spatial domain down-sampling, the problem of global feature loss is avoided; the application is especially suitable for resource-limited scenes such as mobile terminals or edge devices; and the deployment efficiency and practicability of a lightweight model are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing technology, specifically to a deep learning preprocessing method for images based on feature compression. Background Technology

[0002] In the widespread application of deep learning technology, image preprocessing has a critical impact on model performance. Traditional methods often reduce the amount of input data by reducing image resolution to alleviate the problem of high computational resource consumption. However, this spatial domain compression method directly leads to the loss of high-frequency details in the image, making it impossible for the model to fully capture key feature information. Especially when dealing with complex scenes or high-precision tasks, the information loss caused by the reduction in resolution will significantly weaken the model's generalization ability and accuracy. This contradiction is particularly prominent in resource-constrained scenarios such as mobile devices and edge devices. How to retain sufficient feature information while compressing data has become an urgent technical problem to be solved.

[0003] This invention proposes a preprocessing method based on frequency domain feature compression to address the aforementioned problems. By converting the image from the spatial domain to the frequency domain, wavelet transform is used to extract and retain the core visual information in the low-frequency components, while actively discarding high-frequency noise that is insensitive to the human eye. This approach avoids the destruction of global features caused by traditional downsampling and significantly reduces the computational complexity of subsequent deep learning models. The technical solution can flexibly adapt to different scenario requirements by dynamically adjusting the compression strategy, providing a new solution for the efficient deployment of lightweight models. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based image preprocessing method based on feature compression.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a deep learning preprocessing method for images based on feature compression, wherein the preprocessing method includes:

[0006] (1) Frequency domain decomposition: The input image is decomposed into low-frequency components (LL) and high-frequency components (LH, HL, HH) by performing Discrete Wavelet Transform (DWT), where:

[0007] -Low-frequency components are extracted by preserving regions in the wavelet coefficients where the energy percentage is ≥85%;

[0008] - High-frequency components are filtered by energy threshold, retaining local edge information of the top 10% of energy;

[0009] (2) Dynamic parameter adjustment: based on the preset compression ratio formula Dynamically adjust the number of wavelet decomposition levels Based on the semantic category output by the image content analysis module, the corresponding wavelet basis function is selected (Haar wavelet basis function is used for text images, and Daubechies-4 wavelet basis function is used for natural scene images).

[0010] (3) Compression coding: Low-frequency components are input into the deep learning model for feature extraction, while the high-frequency component retention area is sparsely coded (the coding parameters are optimized by the K-SVD algorithm) to ensure that the compressed data volume is ≤ 25% of the original image;

[0011] (4) Reconstructing the approximate original image: Low-frequency components are super-reconstructed using a lightweight generative adversarial network (GAN), and the cross-layer features of multi-scale wavelet coefficients are fused to ensure that the reconstruction error meets the requirements. Furthermore, the texture consistency error is less than 5%.

[0012] As a further aspect of the present invention, the preprocessing method further includes a hybrid frequency domain-spatial domain compression step:

[0013] The image is divided into high-frequency and low-frequency regions. Wavelet compression is used for the high-frequency region, and spatial domain downsampling is used for the low-frequency region. The downsampling ratio is controlled by the following formula:

[0014] ;

[0015] in, To represent the L0 norm of the high-frequency region, that is, the number of non-zero elements in that region, To represent the L0 norm of the entire graph, which is the total number of non-zero elements in the entire graph. Huffman coding optimization of low-frequency components improves storage efficiency by ≥20% after compression.

[0016] As a further aspect of the present invention: the image content analysis module is a lightweight CNN model, which outputs semantic categories including text, faces and natural scenes, with a classification accuracy of ≥95%.

[0017] As a further aspect of the present invention: the preprocessing method includes a dynamic high-frequency supplementation mechanism, which, when the confidence level of the deep learning model... When the value is <0.85, key high-frequency information is recovered through inverse wavelet transform (IDWT) to supplement the ratio.

[0018] As a further aspect of the present invention: the preprocessing method can be extended to the field of video processing, specifically including:

[0019] Frequency domain compression is performed on key frames of the video, and inter-frame motion estimation is used to reduce redundant calculations (compression rate improvement ≥40%).

[0020] When deployed on mobile devices, wavelet transform and feature extraction are processed in parallel using hardware acceleration modules, with a single frame processing time of ≤10ms.

[0021] As a further aspect of the present invention: when the preprocessing method is applied to medical image processing, the gradient amplitude of the lesion edge region in the high-frequency subband is retained (satisfying...). ), and combined with transfer learning to optimize the model’s sensitivity to minor lesions (improving classification accuracy by ≥2%).

[0022] As a further aspect of the present invention: the preprocessing method includes an end-to-end joint optimization mechanism, which embeds a wavelet transform module into a deep learning framework and automatically optimizes compression parameters through backpropagation. The optimization objective function is:

[0023]

[0024] in, The mean square error is between the original image and the reconstructed image.

[0025] As a further aspect of the present invention: in the feature extraction step of the low-frequency components, depthwise separable convolution is used to replace standard convolution (reducing the computational load by 50%-70%), and the core feature weights are enhanced through a channel attention mechanism (improving the model's Top-5 accuracy by ≥3%).

[0026] As a further aspect of the present invention: when deploying the preprocessing method in hardware, a dedicated ASIC chip (with a parallel computing unit optimized for wavelet transform and feature extraction) is designed, which is ≥2 times faster than the traditional GPU and supports real-time processing of 720p video streams (frame rate ≥30fps, power consumption ≤5W).

[0027] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0028] 1. This invention, through frequency domain decomposition and selective compression, significantly reduces the computational burden of deep learning models while preserving the core visual information of images (such as object contours and texture distribution). Compared with traditional spatial domain downsampling, it avoids the problem of global feature loss and is especially suitable for resource-constrained scenarios such as mobile devices or edge devices, significantly improving the deployment efficiency and practicality of lightweight models.

[0029] 2. This invention dynamically adjusts compression parameters (such as wavelet basis functions and decomposition layers) based on image content analysis, supporting the optimization of feature retention range for different semantic categories such as text and natural scenes. Through the controllable discarding and recovery mechanism of high-frequency components, it can not only meet the high-efficiency processing of general tasks, but also flexibly supplement key information for tasks that require details such as target detection, thereby enhancing the versatility and robustness of the technical solution.

[0030] 3. This invention deeply integrates frequency domain compression with deep learning model training, automatically optimizes compression parameters through backpropagation, ensures coordinated adaptation between feature extraction and compression processes, and combines super-resolution reconstruction and high-frequency supplementation techniques to maintain the quality of reconstructed images while reducing data volume, forming a complete closed loop from compression to inference, providing reliable support for application scenarios that emphasize both high precision and high efficiency. Attached Figure Description

[0031] Figure 1 This is a simplified flowchart of a deep learning-based image preprocessing method based on feature compression. Detailed Implementation

[0032] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0033] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0034] Please see the appendix Figure 1 This invention relates to a feature compression-based image deep learning preprocessing method, which includes a preprocessing method comprising:

[0035] (1) Frequency domain decomposition: The input image is decomposed into low-frequency components (LL) and high-frequency components (LH, HL, HH) by performing Discrete Wavelet Transform (DWT), where:

[0036] -Low-frequency components are extracted by preserving regions in the wavelet coefficients where the energy percentage is ≥85%;

[0037] - High-frequency components are filtered by energy threshold, retaining local edge information of the top 10% of energy;

[0038] (2) Dynamic parameter adjustment: based on the preset compression ratio formula Dynamically adjust the number of wavelet decomposition levels Based on the semantic category output by the image content analysis module, the corresponding wavelet basis function is selected (Haar wavelet basis function is used for text images, and Daubechies-4 wavelet basis function is used for natural scene images).

[0039] (3) Compression coding: Low-frequency components are input into the deep learning model for feature extraction, while the high-frequency component retention area is sparsely coded (the coding parameters are optimized by the K-SVD algorithm) to ensure that the compressed data volume is ≤ 25% of the original image;

[0040] (4) Reconstructing the approximate original image: Low-frequency components are super-reconstructed using a lightweight generative adversarial network (GAN), and the cross-layer features of multi-scale wavelet coefficients are fused to ensure that the reconstruction error meets the requirements. Furthermore, the texture consistency error is less than 5%.

[0041] In one embodiment of the present invention, the preprocessing method further includes a hybrid frequency-spatial domain compression step.

[0042] The image is divided into high-frequency and low-frequency regions. Wavelet compression is used for the high-frequency region, and spatial domain downsampling is used for the low-frequency region. The downsampling ratio is controlled by the following formula:

[0043] ;

[0044] in, To represent the L0 norm of the high-frequency region, that is, the number of non-zero elements in that region, To represent the L0 norm of the entire graph, which is the total number of non-zero elements in the entire graph. Huffman coding optimization of low-frequency components improves storage efficiency by ≥20% after compression.

[0045] In one embodiment of the present invention: the image content analysis module is a lightweight CNN model, which outputs semantic categories including text, face and natural scene, with a classification accuracy of ≥95%.

[0046] In one embodiment of the present invention: the preprocessing method includes a dynamic high-frequency supplementation mechanism, which, when the confidence level of the deep learning model... When the value is <0.85, key high-frequency information is recovered through inverse wavelet transform (IDWT) to supplement the ratio.

[0047] In one embodiment of the present invention: the preprocessing method can be extended to the field of video processing, specifically including:

[0048] Frequency domain compression is performed on key frames of the video, and inter-frame motion estimation is used to reduce redundant calculations (compression rate improvement ≥40%).

[0049] When deployed on mobile devices, wavelet transform and feature extraction are processed in parallel using hardware acceleration modules, with a single frame processing time of ≤10ms.

[0050] In one embodiment of the present invention: when the preprocessing method is applied to medical image processing, the gradient amplitude of the lesion edge region in the high-frequency subband is retained (satisfying...). ), and combined with transfer learning to optimize the model’s sensitivity to minor lesions (improving classification accuracy by ≥2%).

[0051] In one embodiment of the present invention: the preprocessing method includes an end-to-end joint optimization mechanism, which embeds a wavelet transform module into a deep learning framework and automatically optimizes compression parameters through backpropagation. The optimization objective function is:

[0052] .

[0053] In one embodiment of the present invention: in the feature extraction step of low-frequency components, depthwise separable convolution is used to replace standard convolution (reducing the computational load by 50%-70%), and the core feature weights are enhanced through channel attention mechanism (improving the model's Top-5 accuracy by ≥3%).

[0054] In one embodiment of the present invention: when the preprocessing method is deployed in hardware, a dedicated ASIC chip (parallel computing unit optimized for wavelet transform and feature extraction) is designed, which is ≥2 times faster than the traditional GPU and supports real-time processing of 720p video streams (frame rate ≥30fps, power consumption ≤5W).

[0055] Example 1: General Image Processing Based on Frequency Domain Feature Compression

[0056] Detailed technical solution:

[0057] 1. Input and Preprocessing:

[0058] Select a 1024×1024 resolution RGB natural scene image (category "coastline") from the ImageNet dataset. The image contains rich texture details (such as waves, rocks, and clouds in the sky). The original image is stored in PNG format and the file size is 2.1MB.

[0059] 2. Frequency Domain Decomposition and Parameter Selection:

[0060] Two-level discrete wavelet transform (DWT) was performed using the Daubechies-4 wavelet basis function. The first level decomposed the wavelet into four subbands: LL1 (512×512), LH1, HL1, and HH1. The second level further decomposed LL1 into LL2 (256×256), LH2, HL2, and HH2.

[0061] The low-frequency component LL2 preserves the main structure of the image (such as the horizon and rock outlines), and its energy proportion is determined by calculating the sum of squares of the wavelet coefficients:

[0062] ;

[0063] In the high-frequency components (LH2, HL2, HH2), the top 10% of local edge information (such as white foam of ocean waves and rock textures) is filtered and retained by setting an energy threshold, while the remaining high-frequency components are set to zero.

[0064] 3. Hybrid compression strategy:

[0065] Spatial domain downsampling: The low-frequency region (flat sky background) is downsampled to 128×128 using bilinear interpolation. The downsampling ratio is controlled by the formula:

[0066] ;

[0067] Based on experimental experience, the settings were set. , At this time, downsampling to a lower resolution is used to improve the compression ratio.

[0068] Sparse coding optimization: The K-SVD algorithm is used to generate a sparse dictionary for the retained high-frequency regions. The dictionary size is 64×256. The sparse coefficients are solved by the OMP (Orthogonal Matching Pursuit) algorithm. The final compressed data size is 23% of the original image (the file size is reduced to 483KB).

[0069] 4. Model Training and Inference:

[0070] The compressed low-frequency components were input into the ResNet-18 model, fine-tuned for 50 epochs on the ImageNet dataset, with a learning rate of 1e-4 and the cross-entropy loss function used.

[0071] Classification performance: Top-5 accuracy reached 82.1%, an improvement of 3.8% compared to the baseline model (78.3% accuracy) which was downsampled to 256×256. The confusion matrix shows that the recall rate of the model for the "coastline" category improved from 75.2% to 81.6%.

[0072] 5. Reconstruction and Quality Verification:

[0073] The image was reconstructed using inverse wavelet transform (IDWT), and a lightweight ESRGAN (1.2M parameters) was introduced for super-resolution enhancement, generating a 1024×1024 resolution image.

[0074] Quantitative indicators:

[0075] PSNR = 32.1 dB (mean square error between original and reconstructed images = 6.4).

[0076] SSIM = 0.963 (error 3.7%), SSIM value of local textures (such as rock surface) ≥ 0.98;

[0077] Visual non-destructiveness was assessed through subjective evaluation by 10 people, with an average Mean Opinion Score (MOS) of 4.6 / 5.0.

[0078] Implementation Results and Comparison:

[0079] Computational efficiency: The GPU (NVIDIA V100) takes 12ms to process a single image, which is 33% more efficient than traditional JPEG2000 compression (18ms).

[0080] Application scenarios: Suitable for scenarios that require a balance between speed and accuracy, such as social media image transmission and autonomous driving environmental perception.

[0081] Example 2: Real-time video stream processing and hardware acceleration

[0082] Detailed technical solution:

[0083] 1. Input and Preprocessing:

[0084] Input a 1080p resolution (1920×1080), 30fps H.264 encoded video stream containing urban road surveillance video (including vehicles, pedestrians, and traffic signs).

[0085] 2. Keyframe frequency domain compression:

[0086] Keyframe selection: Select 1 frame out of every 10 frames as a keyframe (I-frame), and use the Haar wavelet basis function to perform a 1-level decomposition to generate LL1 (960×540), LH1, HL1, and HH1 subbands.

[0087] Dynamic compression strategy:

[0088] The LL1 pixel was downsampled to 480×270 using a bicubic downsampling method, and the compression ratio formula is as follows:

[0089] ;

[0090] High-frequency components retain the top 15% of edge information (such as vehicle outlines and zebra crossings) through energy thresholding, and the data volume is reduced to 28% of the original image after sparse coding.

[0091] 3. Hardware acceleration and parallelization:

[0092] ASIC chip design: Custom 12nm process ASIC, integrating dedicated DWT computing unit (supporting Haar / Daubechies basis functions) and parallel pipeline architecture, with a peak computing power of 16 TOPS.

[0093] Performance metrics:

[0094] The time taken for single-frame wavelet transform has been reduced from 15ms for CPU (Intel i9-12900K) to 3ms for ASIC;

[0095] The overall video stream processing latency is ≤10ms, meeting real-time requirements.

[0096] 4. Inter-frame redundancy elimination:

[0097] Motion estimation: For non-key frames (P frames), optical flow is used to calculate inter-frame motion vectors, reuse the low-frequency components of the previous frame, and only transmit motion residual data, improving the compression rate by 45%.

[0098] Dynamic replenishment mechanism:

[0099] When the target detection model (YOLOv5s) outputs a confidence score < 0.8, proportionally... Restore high-frequency details;

[0100] Experiments show that vehicle detection mAP improved from 74.1% to 76.4% (validated on the COCO dataset).

[0101] 5. Mobile deployment optimization:

[0102] Power consumption control: Deployed on the Snapdragon 8 Gen3 platform, average power consumption is controlled at 4.8W and temperature ≤45℃ through DVFS (Dynamic Voltage Frequency Scaling) technology;

[0103] Real-time performance: Supports 720p@30fps video streaming processing, with end-to-end latency ≤33ms and a stable frame rate of 29.8fps.

[0104] Implementation Results and Comparison:

[0105] Commercial value: Suitable for scenarios such as intelligent traffic monitoring and drone inspection, saving 35% of bandwidth compared to traditional H.265 encoding schemes;

[0106] Robustness test: Under rain and fog interference, the target detection mAP decreased by only 1.2% (compared to 4.5% for the traditional solution).

[0107] Example 3: Optimization of Intelligent Review of Enterprise Contract Documents

[0108] Detailed technical solution:

[0109] 1. Input and Preprocessing:

[0110] Input: A commercial contract in PDF format from a financial institution (containing sensitive clauses, amounts, contracting party information, etc.), with an original file size of 15MB.

[0111] Preprocessing:

[0112] Text content is extracted using OCR (Optical Character Recognition) technology, supporting mixed Chinese and English recognition;

[0113] Wavelet transform compression technology (Daubechies-6 basis functions) is used to reduce noise and compress the scanned image, while preserving the clarity of key text areas (such as signatures, seals, and amounts).

[0114] 2. Key Information Retention and Compression

[0115] Sensitive Information Location:

[0116] Use an NLP model to locate sensitive fields in the contract (such as ID numbers and bank account numbers), and retain their original high-frequency details (compression rate ≤ 30% of the original image).

[0117] Enhanced Error Detection:

[0118] Correct spelling mistakes in the text content (such as miswriting "deposit" as "earnest money"), and verify logical consistency in combination with the rule engine (such as conflicts between amounts and terms).

[0119] 3. Transfer Learning and Model Optimization:

[0120] Pre-training and Fine-tuning:

[0121] Based on the BERT pre-trained model, use 100,000 annotated contract samples for fine-tuning, and focus on optimizing the following tasks:

[0122] Sensitive Information Recognition (F1 value ≥ 95%);

[0123] Clause Compliance Detection (recall rate ≥ 92%).

[0124] Loss Function Design:

[0125] Adopt Focal Loss to alleviate the problem of sample imbalance (such as the sparsity of sensitive fields) and improve the detection ability of small targets.

[0126] 4. Review Process and Efficiency Verification

[0127] Dynamic Supplement Mechanism:

[0128] When the model confidence is < 0.85, automatically trigger the manual review process and highlight the doubtful paragraphs (manual intervention ratio ≤ 5%).

[0129] Performance Metrics:

[0130] The review time is shortened from an average of 25 minutes per copy to 3 minutes per copy;

[0131] The error omission rate is reduced from 8.2% to 2.1% (actual measurement data of a bank customer).

[0132] 5. Enterprise-level Application Effect

[0133] Security:

[0134] Complies with the requirements of GDPR and the Personal Information Protection Law, and supports encrypted storage and audit traceability.

[0135] Commercial Value:

[0136] Saves legal and financial industry clients 90% of manual review costs;

[0137] The contract dispute rate decreased by 40% (data from an insurance company in 2024).

[0138] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Any variations and modifications can be made by those skilled in the art without departing from the spirit and scope of the invention. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention, without departing from the scope of the invention, fall within the protection scope defined by the claims of the present invention.

Claims

1. A deep learning-based image preprocessing method based on feature compression, characterized in that: The preprocessing method includes: (1) Frequency domain decomposition: The input image is decomposed into low-frequency components and high-frequency components by performing discrete wavelet transform, where: -Low-frequency components are extracted by preserving regions in the wavelet coefficients where the energy percentage is ≥85%; - High-frequency components are filtered by energy threshold, retaining local edge information of the top 10% of energy; (2) Dynamic parameter adjustment: based on the preset compression ratio formula Dynamically adjust the number of wavelet decomposition levels And select the corresponding wavelet basis function based on the semantic category output by the image content analysis module; (3) Compression coding: Low-frequency components are input into the deep learning model for feature extraction, while the high-frequency components are sparsely encoded in the retained regions to ensure that the compressed data volume is ≤ 25% of the original image; (4) Reconstructing the approximate original image: Low-frequency components are super-reconstructed using a lightweight adversarial generative network, and the cross-layer features of multi-scale wavelet coefficients are fused to ensure that the reconstruction error meets the requirements. Furthermore, the texture consistency error is less than 5%.

2. The image deep learning preprocessing method based on feature compression according to claim 1, characterized in that: The preprocessing method further includes a hybrid frequency-spatial domain compression step: The image is divided into high-frequency and low-frequency regions. Wavelet compression is used for the high-frequency region, and spatial domain downsampling is used for the low-frequency region. The downsampling ratio is controlled by the following formula: ； in, To represent the L0 norm of the high-frequency region, that is, the number of non-zero elements in that region, To represent the L0 norm of the entire graph, i.e. the total number of non-zero elements in the entire graph, Huffman coding is optimized for low-frequency components, resulting in a storage efficiency improvement of ≥20% after compression.

3. The image deep learning preprocessing method based on feature compression according to claim 2, characterized in that: The image content analysis module is a lightweight CNN model, which outputs semantic categories including text, faces, and natural scenes, with a classification accuracy of ≥95%.

4. The image deep learning preprocessing method based on feature compression according to claim 3, characterized in that: The preprocessing method includes a dynamic high-frequency supplementation mechanism, which adjusts the deep learning model's confidence level as needed. When the value is less than 0.85, key high-frequency information is recovered through inverse wavelet transform, and the ratio is supplemented. .

5. The image deep learning preprocessing method based on feature compression according to claim 4, characterized in that: The preprocessing method can be extended to the field of video processing, and specifically includes: Frequency domain compression is performed on key frames of the video, and inter-frame motion estimation is used to reduce redundant calculations; When deployed on mobile devices, wavelet transform and feature extraction are processed in parallel using hardware acceleration modules, with a single frame processing time of ≤10ms.

6. The image deep learning preprocessing method based on feature compression according to claim 5, characterized in that: When the preprocessing method is applied to medical image processing, it retains the gradient amplitude of the lesion edge region in the high-frequency subband and combines transfer learning to optimize the model's sensitivity to minor lesions.

7. The image deep learning preprocessing method based on feature compression according to claim 6, characterized in that: The preprocessing method includes an end-to-end joint optimization mechanism, which embeds a wavelet transform module into a deep learning framework and automatically optimizes compression parameters through backpropagation. The optimization objective function is: ； in, The mean square error is between the original image and the reconstructed image.

8. The image deep learning preprocessing method based on feature compression according to claim 7, characterized in that: In the feature extraction step of the low-frequency components, depthwise separable convolution is used instead of standard convolution, and the core feature weights are enhanced through a channel attention mechanism.

9. The image deep learning preprocessing method based on feature compression according to claim 8, characterized in that: The preprocessing method uses a dedicated ASIC chip for hardware deployment, which is ≥2 times faster than traditional GPUs and supports real-time processing of 720p video streams.

Images