Industrial part image data enhancement and noise reduction method based on deep learning

Abstract

The invention belongs to the technical field of artificial intelligence and computer vision, and relates to an industrial part image data enhancement and noise reduction method based on deep learning. The method comprises the following steps: step 1, acquiring and preprocessing image data, and constructing a training sample pair required by a data enhancement and noise reduction task; 2, constructing a data enhancement and noise reduction model, wherein the model comprises a double-branch auto-encoder, an encoder feature fusion module, an image reconstruction and decoder module and an image data enhancement module; and 3, inputting the industrial noisy image into the pre-trained data enhancement and noise reduction model, and respectively outputting the noise-reduced and enhanced images. According to the method, tiny edges and flaws can be reserved; the model robustness is improved; and the processing flow is optimized, the calculation load is reduced, the efficient and stable recognition performance is ensured, and the intelligent requirement of industrial manufacturing is met.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and computer vision technology, and relates to a method for enhancing and reducing noise in industrial component image data based on deep learning. Background Art

[0002] In the context of Industry 4.0, AI-powered visual recognition of high-precision parts places higher demands on image enhancement technologies, including noise reduction and data enhancement. Image enhancement, a key means of improving recognition performance, primarily involves noise reduction and data enhancement. Existing image noise reduction methods include traditional image processing techniques such as Gaussian filtering, median filtering, histogram equalization, and the Retinex algorithm. For example, Gaussian filtering smoothes images through convolution to remove noise, histogram equalization enhances contrast by adjusting pixel intensity distribution, and the Retinex algorithm improves low-light image quality by separating illumination and reflection components. However, these methods have significant shortcomings when processing images of high-precision parts: Gaussian and median filtering can blur edges and minor defects, leading to loss of key features; histogram equalization has poor adaptability to complex backgrounds and may introduce artifacts; and the Retinex algorithm struggles with the variable lighting conditions found in industrial scenarios, affecting feature extraction accuracy.

[0003] In terms of data augmentation, traditional methods typically use geometric transformations (such as rotation, flipping, and scaling) or color adjustments (such as brightness and contrast changes) to increase the number of samples to improve the robustness of deep learning models. For example, by rotating an image 0°-180° or scaling it 0.8-1.2 times, different shooting angles and distances can be simulated. However, these simple data augmentation methods struggle to fully simulate the complex variations in high-precision parts in industrial scenarios, such as variable lighting conditions, background interference, or the diversity of small defects, resulting in insufficient generalization capabilities of the model in practical applications. Furthermore, the number of samples of high-precision parts is generally limited, and the sample diversity generated by traditional augmentation methods is limited, making it difficult to meet the requirements of deep learning models for large-scale and diverse data.

[0004] In recent years, deep learning technology has made progress in image enhancement. For example, convolutional neural networks (CNNs) are used for noise reduction, improving clarity by learning image features. For example, the Denoising CNN (DnCNN) removes noise while preserving detail through multi-layer convolution. Generative adversarial networks (GANs) are widely used for data enhancement. Cycle-GAN generates new samples from unpaired data, while Pix2Pix-GAN generates high-resolution images from paired data. Furthermore, literature has proposed frameworks that combine CNNs and GANs, using CNNs to extract features and GANs to generate realistic samples, improving image analysis accuracy. However, the application of these methods in high-precision part recognition still faces challenges: first, CNN noise reduction (such as DnCNN) lacks the ability to preserve subtle defects; second, the quality and diversity of GAN-generated samples (such as Cycle-GAN and Pix2Pix-GAN) are inconsistent, making them difficult to fully meet the high-precision requirements of industrial scenarios. Furthermore, the high computational complexity of deep learning methods makes them difficult to meet the real-time requirements of industry.

[0005] In summary, existing image enhancement technologies face the following challenges in AI visual recognition of high-precision parts: 1) Traditional noise reduction methods (such as Gaussian filtering and Retinex) have poor adaptability to complex scenes and are prone to losing edge and defect features; 2) Traditional data enhancement methods (such as simple geometric transformations) generate insufficient sample diversity, making them incapable of coping with the variability of industrial scenarios; 3) While deep learning methods (such as CNNs and GANs) have improved, their ability to retain minor defects and their real-time performance still require optimization; 4) The limited number of samples limits the training effectiveness of deep learning models, resulting in insufficient recognition accuracy and robustness. Therefore, a method combining efficient noise reduction and diverse data enhancement is urgently needed to improve the image quality and sample diversity of high-precision parts to meet the intelligent needs of industries such as aerospace, aviation, and automotive manufacturing. Summary of the Invention

[0006] This invention aims to address the issues of low image quality, insufficient sample diversity, and poor real-time performance in AI visual recognition of high-precision parts. It addresses feature blurring caused by uneven lighting and noise by providing an efficient noise reduction method that preserves subtle edges and defects. It also develops customized data enhancement strategies to improve model robustness in response to small sample sizes and insufficient diversity. Furthermore, it optimizes the processing flow to reduce computational load and ensure efficient and stable recognition performance to meet the requirements of intelligent industrial manufacturing.

[0007] The present invention provides a method for enhancing and reducing noise of industrial parts image data based on deep learning, comprising the following steps: Step 1: Image data collection and preprocessing to construct training sample pairs required for data enhancement and denoising tasks; Step 2: Build a data enhancement and denoising model, which includes a dual-branch autoencoder, an encoder feature fusion module, an image reconstruction and decoder module, and an image data enhancement module; The dual-branch autoencoder consists of a main branch and a side branch encoder. The main branch is based on PVTv2-B0 and uses a multi-layer multi-head self-attention mechanism to encode the input image using a resolution reduction structure to extract semantic level features. The side branch extracts local detail features through local convolution and downsampling operations. The encoder feature fusion module performs dynamic weighted fusion on the features extracted by the main branch and the side branch; The image reconstruction and decoder module reconstructs the encoded fusion features and outputs a high-quality denoised image; it includes a decoder, a feature-guided attention aggregation module, and a multi-scale channel attention module; The image data enhancement module enhances the image by using geometric transformation, noise simulation and defect simulation based on the denoised image; Step 3: Input the industrial noisy image into the pre-trained data augmentation and denoising model, and output the denoised and enhanced images respectively.

[0008] Preferably, the main branch of the dual-branch autoencoder adopts a resolution reduction structure to perform hierarchical encoding on the input image and extract semantic hierarchical features.

[0009] Preferably, the encoder feature fusion module uses a convolution-transformer attention gate to perform 3×3 convolution and batch normalization on the features extracted by the main branch and the side branch. After fusion, an attention map is generated through ReLU activation and 1×1 convolution to complete the dynamic weighted fusion of global-local features.

[0010] Preferably, the image reconstruction and decoder includes a feature-guided attention aggregation module, which uses three layers of 3×3 convolution and upsampling operations to reconstruct the encoded fusion features.

[0011] Preferably, the image reconstruction and decoder further includes a multi-scale channel attention module, which extracts channel features using convolution kernels of different scales and generates channel attention weights through Sigmoid.

[0012] Preferably, the geometric transformation includes rotating the denoised image by 0°-360°, scaling it by 0.8-1.2 times, and horizontally flipping it to simulate multi-view imaging.

[0013] Preferably, the noise simulation is to add Gaussian noise with a standard deviation range of 0.01 to 0.05 to the denoised image, so as to enhance the robustness of the model to interference.

[0014] Preferably, the defect simulation is to introduce scratches and pit defects into the denoised image to achieve sample enhancement that approximates the real environment.

[0015] Preferably, the Adam optimizer is used to train the model.

[0016] Compared with the prior art, the method of the present invention has the following beneficial effects: (1) Collaborative enhancement of multi-scale features to improve image clarity and feature retention. Through a dual-branch encoder structure, the main branch captures global semantic features based on the Pyramid Vision Transformer, while the side branch uses a lightweight convolutional design to extract local edges and minor defects, achieving full expression of multi-scale features. Experiments show that this mechanism significantly improves image quality in complex industrial scenarios such as uneven lighting and noise interference, with the signal-to-noise ratio increased by approximately 20%. In particular, in high-precision scenarios such as reflective metal surfaces and minor defects, the edge and detail feature retention rate is increased by 15%, providing high-quality input for subsequent AI visual recognition.

[0017] (2) Attention fusion mechanism optimizes noise reduction accuracy. The proposed hierarchical attention module, including the convolution-Transformer attention gate and the feature-guided attention aggregation module, effectively suppresses background noise and highlights key areas by dynamically weighted fusion of global and local features. Experimental results show that the peak signal-to-noise ratio (PSNR) of the denoised image is improved to over 35dB, significantly enhancing image clarity and feature discrimination capabilities compared to Gaussian filtering (approximately 28dB) and DnCNN (approximately 32dB).

[0018] (3) Lightweight design and dynamic supervision are combined to achieve both processing efficiency and robustness. Through lightweight convolutional side branches and the optimized PVTv2-B0 main branch, the model parameters are controlled at approximately 4.38M, the computational load is approximately 0.93GFLOPs, and the single image processing time is ≤0.5 seconds. This significantly improves efficiency compared to traditional CNNs (such as DnCNN, with approximately 15M parameters) and GANs (such as Cycle-GAN, with approximately 10 GFLOPs). It is easy to deploy on industrial cameras or edge devices to meet the real-time detection needs of pipelines.

[0019] (4) Multi-scenario adaptation enhances system stability and reduces industrial deployment costs. In response to lighting changes, perspective diversity, and background interference in industrial scenarios, a customized data augmentation strategy (geometric transformation, noise addition, and defect simulation) generates diverse samples, amplifying the sample size by approximately 8 times, significantly improving the adaptability of upstream and downstream models to complex scenarios. Experiments show that based on the proposed method, the robustness of the downstream recognition model under abnormal conditions such as high noise and blur is improved by approximately 20%. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is an architecture diagram of an industrial parts image data enhancement and noise reduction method based on deep learning in an embodiment of the present invention; Figure 2 Schematic diagram of a convolution-transformer attention gate (CTAG) in an embodiment of the present invention; Figure 3 Schematic diagram of a feature-guided attention aggregation module (FGAAM) in an embodiment of the present invention; Figure 4 Flowchart of a method for industrial parts image data enhancement and noise reduction based on deep learning in an embodiment of the present invention; Figure 5 This is a diagram comparing the image quality of different methods in an embodiment of the present invention; Figure 6 1 is a diagram comparing the efficiency of different models in an embodiment of the present invention; Figure 7 This is a comparison result of downstream recognition robustness under different data enhancement strategies in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] To more clearly understand the technical content of the present invention, its structural scheme is now described in detail with reference to the accompanying drawings and specific embodiments. The accompanying drawings show a preferred embodiment of the present invention, but the scope of protection is not limited thereto. In actual application, it can be flexibly adjusted according to specific usage requirements. This embodiment is intended to help those skilled in the art to fully and deeply understand the present invention.

[0022] The present invention provides a method for industrial parts image data enhancement and denoising based on deep learning (Convolutional Autoencoder, abbreviated as CAE, commonly known in the industry as Convolutional Autoencoder). The method aims to improve image quality and model robustness through efficient noise reduction and customized data enhancement, and meet the real-time recognition needs of industries such as aerospace, aviation, and automobile manufacturing. The architecture of the present invention is as follows: Figure 1As shown in Figure 2, DBAANet employs a four-layer U-shaped encoder-decoder architecture consisting of the following key components: a main-branch encoder based on the Pyramid Vision Transformer (PVT), a lateral convolutional encoder (LCE), and a feature-guided attention aggregation module (FGAAM). The main-branch encoder utilizes the Transformer module to extract global semantic features, while the LCE uses convolution operations to focus on extracting detailed features of the target region, compensating for the Transformer's shortcomings in capturing local features. At the bottleneck layer of the network, DBAANet introduces a convolutional-transformer attention gate (CTAG). This dynamically fuses Transformer and convolutional neural network (CNN) features through an attention mechanism to effectively extract spatially and semantically relevant information. In the decoder, a multi-scale convolutional attention module (MSCAM) is introduced to refine feature maps and enhance the representation of multi-scale contextual information. Furthermore, to further enrich semantic representations, FGAAM dynamically aggregates multi-source features within skip connections, preserving more comprehensive contextual information through an attention mechanism. Finally, the network generates outputs at four different scales, corresponding to the denoised image and three data augmentation samples. These outputs, when fused, support subsequent AI visual recognition.

[0023] Example 1: The method provided by the present invention is as follows Figure 4 The specific steps are as follows: S1: Image data acquisition and preprocessing This step collects images of parts from industrial sites and a general denoising dataset, and performs preprocessing to construct training sample pairs, laying the foundation for subsequent denoising and data augmentation tasks.

[0024] Step 1: Data Collection High-precision image samples of parts containing noise (such as Gaussian noise) and defects (such as scratches and pits) are collected from industrial sites, covering scenarios such as aerospace, aviation, and automotive manufacturing. Noisy images are captured in real-world environments using industrial cameras. The noise comes from ambient lighting, sensor interference, or surface defects. Clear images (target output) are obtained by capturing the same parts using high-precision industrial cameras in controlled environments (such as constant light sources and cleanroom labs), achieving near-noise-free clarity. These images serve as the noise reduction target. Image resolution is typically high (such as 1024×1024 pixels) to capture subtle defect details.

[0025] Step 2: Supplement the general denoising dataset To enhance the model's generalization capabilities, we've supplemented it with general datasets suitable for denoising images of industrial components. These include the PolyU Real Noise Dataset, which contains noisy images from real industrial scenarios (such as the surfaces of mechanical parts), including Gaussian and streak noise, and is close to industrial application environments. The Smartphone Image Denoising Dataset (SIDD) provides image pairs with realistic sensor noise, suitable for simulating the noise characteristics of industrial cameras under complex lighting conditions.

[0026] Step 3: Image Preprocessing Normalize the captured images, standardize the pixel values ​​to the range of [0, 1], and unify the input size (such as 256×256 pixels) to adapt to the model input requirements.

[0027] S2: Dual-branch convolutional autoencoder structure design In this step, a U-shaped convolutional autoencoder structure is designed, which includes a main branch and a side branch encoder to extract global semantics and local detail features respectively.

[0028] Step 1 Main branch encoder design The main branch uses Pyramid Vision Transformer v2 Tiny (PVTv2-B0) as the backbone network and extracts multi-scale global features through a multi-layer multi-head self-attention mechanism (MHSA). Input feature map (size( 、 and )) Through convolutional embedding layers and Transformer block processing, global semantic features are generated to capture the overall structural information of parts.

[0029] Step 2 Side branch encoder design The side branch (LCE) employs a lightweight 3×3 convolutional stack structure, focusing on extracting local details such as edges and minor defects. Unlike traditional multi-layer convolutional architectures, LCE generates feature maps through two convolution operations, combining downsampling and upsampling strategies to achieve multi-scale representation, significantly reducing computational complexity.

[0030] Step 3: Complementary feature extraction The main branch uses MHSA and a feed-forward network (FFN) to model long-range dependencies and capture global semantic information. The side branch extracts local details such as edges and textures through 3×3 convolution, batch normalization, and ReLU activation. The two branches process the input feature map in parallel to generate complementary feature representations.

[0031] S3: Multi-scale feature encoding This step extracts semantic level and local detail features through multi-scale encoding of the main branch and side branches.

[0032] Step 1 Main branch multi-scale encoding The first stage processes the input through the convolutional embedding layer and the Transformer block to generate the first-level feature map , the resolution is reduced to 1 / 4 of the original image. The subsequent stages increase the scaling factor in turn, and the second stage downsampling (scaling factor 2) generates , the third stage generates , the fourth stage generates The Transformer block at each stage contains a multi-head self-attention mechanism (MHSA) and a feed-forward network (FFN) to model long-range dependencies and capture global semantic information such as the overall structure of parts. However, the high scaling factor reduces sensitivity to local details, providing a complementary foundation for side branches.

[0033] Step 2: Side branch local feature extraction Input feature map First, the feature map is generated through the first stage convolution block (including 3×3 convolution, batch normalization BatchNormalization and ReLU activation To obtain deeper features, the second stage convolution is combined with downsampling of a scaling factor of 8 (total scaling factor 32 / 4=8) to generate the fourth level feature map. Intermediate feature map and Through Downsample by 2 and Obtained by 2x upsampling.

[0034] Step 3: Feature complementation The high scaling factor of the main branch may reduce sensitivity to local details, but the side branches compensate for this deficiency through lightweight convolutional design. Features from both branches are transmitted to the decoder via skip connections, enhancing the fusion of global and local information and providing comprehensive input for subsequent feature processing.

[0035] S4: Encoder feature fusion module (CTAG) In industrial parts image processing, a single feature extraction method is difficult to balance global semantic understanding and local detail representation, especially in high-precision edge and defect detection scenarios. To this end, CTAG is introduced in the bottleneck layer of the dual-branch encoder. Its structure is as follows: Figure 2 shown.

[0036] Step 1 Feature alignment and processing CTAG receives the Transformer features of the main branch and LCE characteristics , respectively, through 3×3 standard convolution and batch normalization processing, to extract spatial information and stabilize feature distribution. The specific calculation is: ; .

[0037] Step 2 Joint feature generation Intermediate features are fused through element-wise addition and nonlinearity is introduced through ReLU activation to generate joint features that combine global semantics (such as overall structure) and local details (such as edges and defects). The specific calculation is: .

[0038] Step 3: Attention map generation and weighting An attention map is generated through 1×1 convolution, batch normalization, and sigmoid activation, indicating the importance of each spatial location in the feature map. The attention map weights the input features and generates dynamically fused output features, improving the processing capability of complex scenes. The specific calculation is: ; ; in, and They are 3×3 and 1×1 convolution operations respectively, is batch normalization, is the rectified linear unit activation function, Generate an attention map in the range [0, 1], is the attention weight matrix, is the final output feature matrix of CTAG.

[0039] S5: Image reconstruction and decoder structure design This step reconstructs a high-quality denoised image through the decoder and introduces FGAAM to optimize the multi-scale feature expression. The specific structure is as follows: Figure 3 shown.

[0040] Step 1 Decoder Structure The decoder uses three layers of 3×3 convolution and upsampling modules to gradually restore the encoded fusion features to the original resolution and generate a denoised image. Upsampling is combined with skip connections to fuse the multi-scale features in the encoder, such as and ( ). The specific calculation is: ; ; .

[0041] Step 2: FGAAM feature aggregation To efficiently fuse multi-scale features between the encoder and decoder during the decoding phase, we propose FGAAM. FGAAM progressively aggregates feature maps through an attention mechanism, leveraging the network's learned attention coefficients to enhance the activation of relevant features and suppress irrelevant information. FGAAM receives feature maps from the encoder's main branch and LCE, as well as upsampled features from the deep decoder's MSCAM. It generates intermediate features through 3×3 grouped convolution and batch normalization. These intermediate features are fused through element-wise addition and ReLU activation to generate aggregated features. The specific calculation is as follows: .

[0042] Step 3: Attention-guided output Attention guidance generates attention coefficients for three channels through 1×1 convolution and sigmoid activation, weighting the decoder, LCE, and main branch features respectively to generate the final attention map, and outputs the attention-guided aggregated features through element-wise multiplication. FGAAM's multi-feature guidance architecture captures both fine-grained and coarse-grained semantic information, optimizes computational efficiency, and enhances denoising and enhancement capabilities for industrial images. The specific calculation is as follows: Attention coefficient generation: ; .

[0043] Final attention map: .

[0044] Attention-guided output: ; in, is a 3×3 group convolution, Split the feature map into three channels, 、 、 are the attention coefficients of the decoder, LCE and main branch respectively, is the final attention map, is the output feature of FGAAM.

[0045] S6: Introduction of Multi-Scale Channel Attention Module (MSCAM) Following FGAAM, MSCAM further refines feature maps and enhances multi-scale contextual information. It extracts channel features using multi-scale convolution kernels (3×3, 5×5) and applies sigmoid activation to generate channel attention weights, optimizing feature representation and improving the robustness of noise reduction and data augmentation.

[0046] S7: Industrial Image Noise Reduction Processing Step 1: The decoder of the present invention processes features by multi-layer downsampling (1 / 4, 1 / 8, 1 / 16, 1 / 32 resolution) and fusion of attention of CTAG, FGAAM, and MSCAM, and finally generates up 4× HEAD .

[0047] Step 2: Noisy image After multi-layer downsampling by a dual-branch encoder (main branch based on PVTv2-B0 and LCE side branch), CTAG fuses global and local features at the bottleneck layer.

[0048] Step 3: FGAAM and MSCAM gradually upsample in the decoder stage, integrate multi-scale information, and reconstruct As the noise reduction output. Experiments show The peak signal-to-noise ratio (PSNR) reaches 35dB, the signal-to-noise ratio is improved by 20%, and the processing time is 0.45 seconds.

[0049] S8: Design of customized image data enhancement strategies Data enhancement processing (output 、 、 ).

[0050] (1) Geometric transformation (output ): Step 1: By Processing and fusing previous lower-level features ( and Fusion, corresponding to 1 / 8 resolution), processed by the geometric transformation module .

[0051] Step 2: As the initial denoised image, it is integrated with the previous lower layer feature map (containing edge and texture information) through a feature fusion layer (such as 3×3 convolution and batch normalization). Next, random rotation (covering 0° to 360°, step size 15°), scaling (range 0.8 to 1.2 times, step size 0.1) and horizontal / vertical flipping (probability 50%) are applied to generate diverse perspective samples. After processing, the output .

[0052] Step 3: .in is the geometric transformation function, is the previous lower-level feature, is the rotation angle, is the scaling factor, is the flip flag (0 or 1). About 300 variant images are generated, covering multiple perspectives and scale changes, enhancing the model's adaptability to different shooting angles. Experiments show that the robustness is improved by about 15% and the sample diversity increases by 20% compared to when the lower-level features are not integrated.

[0053] (2) Noise addition (output ): Step 4: Generated by up 16× HEAD, fusion of previous lower layer features (such as or , corresponding to 1 / 16 resolution), processed by the noise adding module .

[0054] Step 5: Compared with the previous lower layer feature map (including mid-scale structural information) is integrated through the feature fusion layer to extract local texture details. Then, Gaussian noise is superimposed, and the standard deviation Random sampling between 0.01 and 0.05 (step size 0.01) to simulate lighting or sensor noise in industrial environments. Up to 16× HEAD output .

[0055] Step 6: ,in is the Gaussian noise function, is the previous lower layer feature, is the noise variance: About 200 noisy variant images were generated, and the noise range covered common industrial interference. After fusing the lower-level features, the detail retention rate was improved by about 18%, and the accuracy of the subsequent model in low signal-to-noise ratio scenarios was improved to 94%.

[0056] (3) Defect simulation (output ): Step 7: Generated by up 32× HEAD, fusion of previous lower layer features (such as or , corresponding to 1 / 32 resolution), processed by the defect simulation module .

[0057] Step 8: Compared with the previous lower layer feature map The feature fusion layer extracts high-level defect patterns (including deep semantic information). Next, scratches (5-15 pixels in length, 1-3 pixels in width) and pits (3-7 pixels in diameter) are randomly added with a uniform density between 0.01 and 0.05 to simulate surface defects. (Up to 32× HEAD output) .

[0058] Step 9: ,in is the defect function, is the previous lower layer feature, is the density, is the scratch parameter, is the concave point parameter. About 250 defect variant images were generated, and the defect distribution met industrial standards. After fusing the underlying features, the defect detection rate increased by about 20%, and the model's recognition accuracy for complex defects reached 95%.

[0059] S9 training optimization strategy and model convergence This step optimizes model training to ensure efficient convergence and low computational overhead.

[0060] Step 1 Optimizer Settings The Adam optimizer (initial learning rate 0.001, weight decay 0.0001) was used, and the mean square error (MSE) was used as the loss function. The model parameters were optimized after 100 training rounds.

[0061] Step 2: Model Lightweighting Through a dual-branch lightweight design and an efficient attention mechanism, the model parameters are controlled at approximately 4.38M, the computational overhead is approximately 0.93 GFLOPs, and the processing time for a single image is less than 0.5 seconds.

[0062] Step 3: Convergence Effect Dynamic supervision and lightweight design ensure fast model convergence. Experiments show that the loss is stable after training, and the denoising and enhancement performance are better than the traditional U-Net model.

[0063] Example 2: The present invention designs relevant experiments to verify the effectiveness of the model and method proposed in the present invention. The specific contents are as follows: Experiment 1: Image quality comparison experiment Experimental methods: Images of industrial parts containing noise, reflective metal surfaces, and subtle defects were selected as test data. Gaussian filtering, median filtering, Cycle-GAN (cyclic generative adversarial network), DnCNN (denoising convolutional neural network), and the proposed method were used for noise reduction. Using the noise-free original image as a reference, the peak signal-to-noise ratio (PSNR) of each method was calculated. The gradient-based edge detail preservation index (EPI_weighted) was also used to measure the fidelity of the image at edges and in areas with minor defects.

[0064] Experimental steps: Step 1: Collect and label a batch of high-precision parts images and construct a noisy version as input.

[0065] Step 2: Use Gaussian filtering, median filtering, Cycle-GAN, DnCNN and the method of the present invention to perform image restoration respectively.

[0066] Step 3: Calculate the PSNR of the denoised image and the original noise-free image.

[0067] Step 4: Extract edge features (Sobel / Laplacian operator), construct weighted masks in the reflective and high gradient areas, and calculate EPI_weighted. Summarize the results and draw a comparison chart, such as Figure 5 shown.

[0068] Experimental results: Results show that the proposed method achieves a PSNR of approximately 35 dB, significantly higher than Gaussian filtering (≈28 dB), median filtering (≈27 dB), Cycle-GAN (≈29 dB), and DnCNN (≈32 dB). In terms of the EPI_weighted metric, the proposed method achieves an average improvement of approximately 15% over other methods, preserving more edge and detail information on reflective metal surfaces and subtle defects.

[0069] Experiment 2: Efficiency comparison experiment To verify the efficiency advantage of the method of the present invention, DnCNN, Cycle-GAN and the method of the present invention were selected for comparison. The number of parameters (M) and the amount of computation (GFLOPs) of each model were counted. GFLOPs scales linearly with the input size. This experiment was uniformly evaluated on 256×256.

[0070] The specific experimental steps are as follows: Step 1: On the same hardware platform (RTX 3090 / i7-12700F; PyTorch 2.0), load DnCNN, Cycle-GAN, and the proposed method respectively.

[0071] Step 2: Count the number of model parameters (M) and the amount of computation for a single forward propagation (GFLOPs).

[0072] Step 3: Compare the scale and computational efficiency of different models. Draw a double bar chart with the model (method) name on the horizontal axis. Each model shows two indicators: parameter count and GFLOPs, such as Figure 6 shown.

[0073] Experimental results show that the proposed model has approximately 4.38M parameters and 0.93 GFLOPs of computation, significantly lower than DnCNN and Cycle-GAN. This demonstrates that the proposed method is not only lightweight but also computationally efficient, making it more suitable for deployment on industrial cameras or edge devices, meeting the real-time detection requirements of production lines.

[0074] Experiment 3: Downstream recognition robustness experiment To evaluate the impact of different data augmentation strategies on downstream recognition models, we use ResNet50 as the classifier and compare the following training data schemes under the same training and test split conditions: a. Raw data; b. Traditional geometric enhancement data (rotation, cropping, etc.); c. Cycle-GAN augmented data; d. Enhanced data generated by the model of the present invention (noise reduction preprocessing + customized diversified enhancement).

[0075] Step 2: During the test phase, evaluate the accuracy under clean data (Clean) and perturbed data (Perturbed, which contains noise, blur, lighting, and slight viewpoint changes), and calculate the accuracy drop Δ.

[0076] The specific experimental steps are as follows: Step 1: Divide the same industrial parts dataset into training set and test set.

[0077] Step 2: Use the above four types of training data schemes to train the ResNet50 classifier respectively.

[0078] Step 3: Record the accuracy Acc_clean on the Clean test set and Acc_perturbed on the Perturbed test set.

[0079] Step 4: Calculate the robustness improvement percentage. Robustness improvement is defined as R = (Δ_baseline - Δ_method) / Δ_baseline × 100%, where Baseline is the original data training scheme, Δ_baseline is the difference between Acc_clean and Acc_perturbed under the original data, and Δ_method is the difference between Acc_clean and Acc_perturbed under a certain model. Draw a grouped bar chart to show the accuracy of each model under Clean and Perturbed, and draw a broken line to mark the robustness improvement percentage, such as Figure 7 shown.

[0080] Results show that the baseline method suffers a significant drop in accuracy under perturbations. While traditional geometric augmentation can partially mitigate this, the drop remains significant (robustness improved by approximately 11%). Cycle-GAN augmentation performs even worse under noise and viewpoint perturbations, with robustness decreasing by approximately 4% compared to the baseline. Using training samples generated using our method, the model exhibits the smallest drop in accuracy on the Perturbed test set, improving overall robustness by approximately 20% compared to the baseline. These results demonstrate that our method can significantly enhance the stability of downstream recognition models in complex industrial scenarios.

Claims

1. A method for industrial parts image data enhancement and denoising based on deep learning, characterized in that: The following steps are involved: Step 1: Image data collection and preprocessing to construct training sample pairs required for data enhancement and denoising tasks; Step 2: Build a data enhancement and denoising model, which includes a dual-branch autoencoder, an encoder feature fusion module, an image reconstruction and decoder module, and an image data enhancement module; The dual-branch autoencoder includes a main branch and a side branch encoder. The main branch is based on PVTv2-B0 and uses a multi-layer multi-head self-attention mechanism to encode the input image using a resolution reduction structure to extract semantic level features. The side branch extracts local detail features through local convolution and downsampling operations; The encoder feature fusion module performs dynamic weighted fusion on the features extracted by the main branch and the side branch; The image reconstruction and decoder module reconstructs the encoded fusion features and outputs a high-quality denoised image; it includes a decoder, a feature-guided attention aggregation module, and a multi-scale channel attention module; The image data enhancement module enhances the image by using geometric transformation, noise simulation and defect simulation based on the denoised image; Step 3: Input the industrial noisy image into the pre-trained data augmentation and denoising model, and output the denoised and enhanced images respectively.

2. The method for industrial parts image data enhancement and noise reduction based on deep learning according to claim 1 is characterized in that: The main branch of the dual-branch autoencoder uses a resolution reduction structure to hierarchically encode the input image and extract semantic hierarchical features.

3. The method for industrial parts image data enhancement and noise reduction based on deep learning according to claim 1 is characterized in that: The encoder feature fusion module uses a convolution-transformer attention gate to perform 3×3 convolution and batch normalization on the features extracted from the main branch and the side branch. After fusion, an attention map is generated through ReLU activation and 1×1 convolution to complete the dynamic weighted fusion of global-local features.

4. The method for industrial parts image data enhancement and noise reduction based on deep learning according to claim 1, characterized in that: The feature-guided attention aggregation module receives feature maps from the main branch and side branches of the encoder, as well as upsampled features from the multi-scale channel attention module, and generates intermediate features through 3×3 grouped convolution and batch normalization; The intermediate features are fused through element-wise addition and ReLU activation to generate aggregated features.

5. The method for industrial parts image data enhancement and noise reduction based on deep learning according to claim 1, characterized in that: The multi-scale channel attention module uses convolution kernels of different scales to extract channel features and generates channel attention weights through Sigmoid.

6. The method for industrial parts image data enhancement and noise reduction based on deep learning according to claim 1, characterized in that: The geometric transformation includes rotating the denoised image by 0°-360°, scaling it by 0.8-1.2 times, and horizontally flipping it to simulate multi-view imaging.

7. The method for industrial parts image data enhancement and noise reduction based on deep learning according to claim 1, characterized in that: The noise simulation is to add Gaussian noise with a standard deviation range of 0.01 to 0.05 to the denoised image to enhance the robustness of the model to interference.

8. The method for industrial parts image data enhancement and noise reduction based on deep learning according to claim 1, characterized in that: The defect simulation is to introduce scratches and pit defects into the denoised image to achieve sample enhancement that is close to the real environment.

9. The method for industrial component image data enhancement and noise reduction based on deep learning according to any one of claims 1 to 8, characterized in that: The Adam optimizer is used to train the model.

Images