A real-time all-in-one image enhancement method, electronic device and storage medium

Abstract

This invention discloses a real-time multi-functional image enhancement method, electronic device, and storage medium, belonging to the field of deep learning image restoration technology. The method includes: acquiring a degraded image to be processed; using an encoder to downsample the degraded image to generate a first latent feature; using a discriminant network to process the degraded image to generate a second latent feature representing the degradation type information of the degraded image; modulating the first latent feature based on the second latent feature to generate a fused feature that integrates degradation type information; and using a decoder to upsample the fused feature to generate a restored image. This invention automatically identifies degradation types through a discriminant network, can automatically and dynamically perceive the degradation components of the input image, and simultaneously modulates the first latent feature based on the second latent feature, achieving accurate balanced restoration of scenes with multiple degradation factors superimposed.

Description

Technical Field

[0001] This invention relates to the field of deep learning image restoration, and more particularly to a real-time all-in-one image enhancement method, electronic device, and storage medium. Background Technology

[0002] Deep learning has been widely applied in low-quality image enhancement (LQIE) across various real-world scenarios, such as nighttime surveillance, autonomous driving, medical imaging, underwater exploration, and military reconnaissance. Low-quality images and videos typically suffer from severe noise, blurred details, and low dynamic range, making it difficult for traditional enhancement methods (such as histogram equalization or multi-frame averaging) to effectively recover visible information.

[0003] All-in-One Image Restoration (AiOIR) deep learning models integrate deblurring, super-resolution, low-light enhancement, and dehazing functions, demonstrating unique value in multiple fields. Compared to single-task models, its core advantages are significant: first, improved efficiency by avoiding redundant computations in multi-model cascading; second, optimized resources by reducing hardware deployment costs; and third, synergistic gains, as multi-task joint optimization can leverage cross-task feature associations to improve overall restoration performance in complex scenes.

[0004] However, the following challenges still exist in practical applications:

[0005] 1. Poor adaptability to complex scenarios: When multiple degradation factors overlap, it is difficult to balance the priorities of multiple tasks;

[0006] 2. Feature conflicts between tasks: Feature conflicts exist between different tasks, making joint optimization difficult and easily leading to situations where the performance of one task is improved while the performance of other tasks degrades.

[0007] 3. Insufficient generalization ability: The training data is difficult to cover all real-world scenarios, and it is prone to failure in unseen degradation modes. Summary of the Invention

[0008] One of the objectives of this invention is to provide a real-time all-in-one image enhancement method, electronic device, and storage medium to solve the aforementioned problems.

[0009] This invention is achieved through the following technical solution: a real-time multi-in-one image enhancement method, the image enhancement method comprising:

[0010] Obtain the degraded image to be processed;

[0011] An encoder is used to downsample the degraded image to generate a first latent feature;

[0012] A discriminative network is used to process the degraded image to generate a second latent feature that characterizes the degradation type information of the degraded image;

[0013] The first latent feature is modulated based on the second latent feature to generate a fused feature that incorporates degradation type information;

[0014] A decoder is used to upsample the fused features to generate the restored image.

[0015] Further, the modulation of the first latent feature based on the second latent feature includes:

[0016] The first latent feature and the second latent feature are multiplied element-wise to obtain the fused feature.

[0017] Furthermore, the modulation of the first latent feature based on the second latent feature further includes:

[0018] The features obtained through the element-wise multiplication operation are then combined with the first latent features using a weighted combination.

[0019] Furthermore, the encoder and the decoder constitute a U-Net framework, and

[0020] The decoder is configured to, during upsampling, aggregate the intermediate features with the features generated by the encoder during downsampling, which correspond to the size of the intermediate features.

[0021] Furthermore, the downsampling structure of the discriminant network and the downsampling structure of the encoder are constructed using the same modules.

[0022] Furthermore, the downsampling process in the encoder and / or the discriminant network includes at least a basic module employing the following steps:

[0023] The input features are processed through the first convolutional layer;

[0024] The output of the first convolutional layer is split along the channel dimension to obtain the first branch features and the second branch features; channel attention weighting is applied to the first branch features;

[0025] The first branch features and the second branch features, which are weighted by channel attention, are aggregated to generate aggregated features;

[0026] as well as

[0027] The aggregated features are subjected to a block multiplication operation to achieve feature filtering.

[0028] Furthermore, during the training phase, the discriminative network includes a classification head and is trained using a classification loss function to enable it to acquire degenerate recognition capabilities; during the inference phase, the classification head is removed.

[0029] In another aspect, the present invention provides an electronic device including a processor and a memory, wherein the memory stores a computer program, and the processor is configured to, when executing the computer program, implement any one of the above-described real-time multi-image enhancement methods.

[0030] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the method described in any one of the above-described real-time multi-image enhancement methods.

[0031] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0032] Traditional or single-model approaches often fail to perform well when processing images with unknown degradation types due to a lack of prior information. This invention utilizes a discriminant network to automatically identify degradation types, enabling automatic and dynamic perception of degradation components in the input image. Without manual prompting, it can adaptively identify and weight mixed degradation factors in the image, such as simultaneously handling blur, low light, and noise. This achieves balanced restoration of complex scenes and avoids the problem of over-optimization of a single task leading to degradation in other tasks.

[0033] This invention modulates the first latent feature based on the second latent feature, dynamically adjusting the recovery strategy. The model no longer requires fixed, rigid task priorities. For degradation-related parts of the first latent feature, it automatically suppresses degradation-related features and enhances features related to clear content. Through this pixel-level fine-tuning, it can explicitly focus on and process the most severe degradation problems in the current image, thereby achieving noise reduction without sacrificing edge sharpness and maintaining good color fidelity while dehazing. It achieves accurate balanced recovery in scenes with multiple degradation factors, solves the feature conflict problem between multiple tasks, and realizes cross-task feature collaboration. Attached Figure Description

[0034] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0035] Figure 1 This is a schematic diagram of a real-time multi-image enhancement method provided by the present invention.

[0036] Figure 2 This is a schematic diagram of the structure of a real-time multi-in-one image enhancement method provided by the present invention.

[0037] Figure 3 This is a partial structural diagram of the downsampling module in this invention.

[0038] Figure 4 This is a schematic diagram of the element-wise product fusion structure of multi-scale features and discriminative features in this invention.

[0039] Figure 5 This is a schematic diagram of the model flow structure in this invention.

[0040] Figure 6 This is a diagram showing the deblurring effect in this invention.

[0041] Figure 7 This is a diagram showing the defogging effect in this invention.

[0042] Figure 8 This is a diagram illustrating the effect of enhancing low-light performance in this invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0044] Example 1

[0045] Deep learning-based image restoration has made significant progress in single-task restoration of various degradation types, demonstrating its ability to reconstruct high-quality image structures when faced with degraded image inputs from a single domain. Decomposing complex real-world images with mixed domains (such as images with low light, blur, and noise due to focus failure at night) into multiple single-task degradations is a suboptimal approach. In fact, the objective function for multi-degrade mixed image restoration is consistent, implying a cross-correlation between degradation types. The most critical challenge in multi-degrade image restoration tasks lies in informing the image restoration model of the degradation components in the input image. Currently, state-of-the-art multi-degrade methods all employ incorporating prompts into the model to inform it of the degradation components of the input image. This requires contrastive learning models or manual input, which does not meet the performance and automation requirements for practical industrial applications.

[0046] However, multi-task image restoration models face multiple challenges in practical applications. First, there's the issue of adaptability to complex scenes. Real-world environments often involve multiple degradation factors, such as blurriness, low light, and noise simultaneously in rainy conditions. Models must balance the priorities of multiple tasks, easily leading to neglecting some tasks for others. Second, there's the conflict between real-time performance and overall performance. Multi-task integration increases the computational load of individual models, making it difficult to balance processing speed and restoration effectiveness on computationally limited devices like drones and vehicle systems, potentially causing decision delays. Third, there's insufficient generalization ability. Training data struggles to cover all real-world scenarios, such as specific fog concentrations or extreme low-light environments, making models prone to failure under unseen degradation patterns. Furthermore, feature conflicts exist between different tasks; for example, super-resolution may amplify noise, making joint optimization difficult and prone to situations where performance improvement in one task leads to degradation in others, impacting overall reliability. To address the shortcomings of current multi-faceted image restoration methods, this invention provides a multi-faceted, all-weather adaptive image restoration method. Unlike existing methods that employ cue learning, contrastive learning, or multimodal information fusion, this invention uses only lightweight convolutional structures to achieve multi-faceted restoration, offering simplicity, efficiency, high real-time performance, and edge deployment capabilities. Specifically, a fast restoration U-Net framework is built using efficient convolutional modules. A lightweight discriminant network extracts the latent category semantics of the image. The latent feature vectors of the U-Net framework are then pixel-level filtered based on the class scores of the discriminant network. Since the scores significantly change pixel weights, this enables the model to explicitly detect degradation factors, allowing for cross-task feature association during the restoration phase, resulting in a high-quality image. Extensive experiments have shown that this discriminant score-weighted method achieves performance comparable to currently popular cue learning fusion methods, fully demonstrating the effectiveness of this invention.

[0047] This embodiment discloses a real-time multi-in-one image enhancement method. Figure 1 A flowchart of a real-time all-in-one image enhancement method in this embodiment is shown. As can be seen from the figure, this embodiment includes the following steps:

[0048] Step 1: Obtain the degraded image to be processed;

[0049] First, raw, low-quality image data is received from a specified data source (such as a camera sensor, storage medium, or network stream). This image is typically in RGB three-channel format, and its resolution can vary depending on the application scenario, for example, 512x512 pixels on mobile devices or 256x256 pixels in a real-time video stream. Before being input into the network, the image data undergoes normalization preprocessing, normalizing pixel values ​​from the integer range [0,255] to the floating-point range [0,1] or [-1,1] to accelerate model training convergence and improve numerical stability.

[0050] Step 2: Use an encoder to downsample the degraded image to generate a first latent feature;

[0051] The encoder consists of multiple downsampling modules used to sequentially execute the following steps:

[0052] Convolutional processing: The input features are processed through a first convolutional layer, which can be set with a stride to achieve downsampling or maintain size as needed.

[0053] Channel splitting: The convolution output is normalized and then split in half along the channel dimension into first branch features and second branch features.

[0054] Channel attention weighting: A lightweight channel attention mechanism is applied to the features of the first branch. Channel weights are generated by global average pooling and multiplied to achieve adaptive calibration for different channels.

[0055] Feature aggregation: The first branch features that have undergone channel attention weighting are superimposed and fused with the unweighted second branch features to generate aggregated features.

[0056] Feature Filtering: Further filtering and compression of aggregated features. First, pointwise convolution is used to stretch the channel dimension. Then, the stretched features are split into at least two sub-feature blocks. These sub-feature blocks are multiplied element-wise. Finally, a pointwise convolution is used to weight the multiplication result by the channel dimension, completing the final feature filtering and output.

[0057] After several such cascaded downsampling modules, the image size gradually decreases while the number of channels gradually increases. Finally, the encoder outputs a highly abstract, semantically rich first latent feature at the bottom layer, with typical dimensions of [64, 64, 128]. This feature comprehensively encodes the content structure information of the degraded image.

[0058] Step 3: A discriminative network is used to process the degraded image to generate a second latent feature that characterizes the degradation type information of the degraded image;

[0059] During model inference, the encoder extracts features from the input image and outputs multi-scale first latent features. In parallel, a lightweight discriminative network processes the same input image and outputs second latent features to characterize degradation type information.

[0060] The lightweight discriminative network does not possess a complete downsampling structure independent of the encoder. Instead, its downsampling part is constructed using the exact same downsampling modules as the encoder's downsampling part. In practice, this means that the input image first undergoes feature extraction through a shared set of basic downsampling modules, and the resulting intermediate features are then input into subsequent parts of the encoder and the discriminative network for processing. This structure-sharing mechanism significantly reduces the overall number of parameters and computational cost of the model. It avoids the computational redundancy caused by designing a separate feature extractor for the discriminative network, allowing the model to increase its powerful degradation recognition capabilities without incurring additional computational overhead. This enables the invention to be efficiently applied to edge devices with limited computing power.

[0061] This invention constructs a U-Net based on a lightweight convolutional module as the main recovery network and introduces a lightweight discriminant network. Unlike traditional Generative Adversarial Networks (GANs) where the discriminator is only used for real / false judgment, this lightweight discriminant network's core function is to learn the degradation features of the input image. Before using the discriminant network, it needs to be trained. Specifically, a classification head is connected to the output of the discriminant network. This classification head typically consists of a global pooling layer and a multilayer perceptron (MLP) to compress the features extracted by the discriminant network into a vector of the same dimension as the preset number of degradation types. Subsequently, the discriminant network is trained using a cross-entropy loss function to supervise its learning to recognize the degradation components of the input image, such as determining whether the image is blurry, foggy, or dark. After the model completes training, the classification head is removed. At this point, the discriminant network no longer outputs specific class probabilities but instead outputs a second latent feature rich in degradation semantic information generated by its last convolutional layer. This feature is used for modulation and fusion with the first latent feature in the encoder.

[0062] This decoupled training-inference design allows the discriminative network to efficiently learn representations of degradation features during training through explicit classification tasks, while in the deployment phase it becomes a lightweight, manual-free degradation feature extractor. It achieves fully automated degradation perception while avoiding unnecessary classification calculations during inference. Furthermore, the discriminative network learns the essential features of various degradation types during training. Even when encountering a mixed degradation not present in the training set at the exact same proportion—for example, a specific type of fog with uneven concentration—the discriminative network can identify it as a variant related to fog based on its learned features and output an approximate discriminative feature. This solves the problem that training data cannot cover all real-world scenarios, such as specific fog concentrations or extreme low-light environments, and is prone to failure in unseen degradation patterns.

[0063] Step 4: Modulate the first latent feature based on the second latent feature to generate a fused feature that incorporates degradation type information;

[0064] After obtaining the second and first latent features, the first and second latent features are multiplied element-wise to obtain the fused features. This operation is not a simple feature concatenation, but a fine-grained, pixel-level dynamic modulation mechanism. The second latent feature can adaptively scale different spatial locations and channels in the first latent feature. For feature activations associated with severe degradation, this operation has a suppressive effect; for feature activations associated with clear content, it has an enhancing effect. This enables the restoration network to detect degradation factors, thereby achieving adaptive and targeted image restoration and resolving the feature conflict problem between multiple tasks.

[0065] The initial fused features are then weighted and summed with the original first latent features output by the encoder using learnable parameters to generate the final fused features. This ensures that while introducing discriminative information for modulation, the content information of the original image is preserved to the greatest extent possible, preventing excessive distortion or loss of key content features during the modulation process. Through the weighting of learnable parameters, the model can adaptively balance the ratio between dependence on discriminative priors and reliance on original content, thereby generating more stable and reliable fused features and improving the quality and naturalness of image reconstruction.

[0066] In real-world scenarios, multiple degradation factors (such as rain / fog + blur + low light) often overlap. Models need to balance the priorities of different tasks. This invention modulates the first latent feature based on the second latent feature, dynamically adjusting the recovery strategy. The model no longer needs fixed, rigid task priorities. For degradation-related parts of the first latent feature, such as low-frequency information corresponding to blur and abnormal activation values ​​corresponding to noise, the discriminative feature will suppress them; for parts related to clear content, they will be enhanced. Through this pixel-level fine-tuning, the most severe degradation problem in the current image can be explicitly addressed and processed, thus achieving denoising without sacrificing edge sharpness and maintaining good color fidelity while dehazing. This achieves precise balanced recovery of scenes with multiple degradation factors overlapping, solves the feature conflict problem between multiple tasks, and realizes cross-task feature collaboration.

[0067] Step 5: Use a decoder to upsample the fused features to generate the restored image.

[0068] The encoder and decoder together constitute a U-Net image restoration network. The encoder consists of multiple downsampling modules. The upsampling decoder needs to fully reconstruct the encoded features that incorporate discriminative information. Considering the efficiency of edge deployment, it is also constructed using downsampling modules. However, unlike downsampling, upsampling requires amplifying the feature image and reducing the channel dimension. In current popular practices, pixel shuffle is generally used, but this requires a 4x amplification of the channel dimension during module processing, which is computationally expensive. In this invention, upsampling interpolation can achieve sufficiently good restoration results. Specifically, a convolutional layer compresses the number of channels while maintaining the size, followed by bilinear interpolation to amplify the feature map size. Compared to the computationally intensive pixel reassembly method, this choice is more conducive to edge deployment. After upsampling at each layer, the decoder aggregates information with the features at the corresponding scale of the encoder's intermediate layers (usually a feature addition operation). For example, the features upsampled to [128, 128, 64] are added to the [128, 128, 64] features generated during the encoder's downsampling process. This mechanism effectively transmits the low-level details (such as edges and textures) captured by the encoder, compensates for the information loss that may be caused by upsampling, and ensures the detail fidelity of the reconstructed image. After several upsampling and feature fusion, the feature map size is restored to the original input size (e.g., [512,512,16]), and finally the number of channels is compressed to 3 through a 3x3 convolutional layer, outputting the final restored RGB image [512,512,3].

[0069] U-Net's symmetrical structure and skip connection mechanism enable the rich low-level features (such as edges and textures) captured by the encoder to be directly passed to the corresponding layers of the decoder. This effectively compensates for the information loss caused by the network's downsampling and upsampling operations, providing the decoder with accurate detail information for high-quality image reconstruction, thus ensuring the fidelity of the output image in terms of edge sharpness and texture detail.

[0070] Application examples

[0071] The following combination Figures 3-5 The invention will be further illustrated with specific applications, but it should be noted that actual operation is not limited to the data dimensions in this example. This case study demonstrates the complete process and effects of this method by implementing it on a device equipped with a mobile SoC.

[0072] I. Hardware and Software Environment:

[0073] Hardware platform: Mobile devices equipped with Snapdragon 8gen1 SoC.

[0074] Software environment: Deep learning frameworks such as PyTorch or TensorFlow Lite are used for model inference.

[0075] Input image: A 512x512 resolution RGB degraded image, such as an image taken at night with motion blur and significant noise.

[0076] II. Model Structure and Data Processing Flow:

[0077] The model used in this case study mainly consists of an encoder with a shared downsampling module, a lightweight discriminant network, and a decoder.

[0078] Input and preprocessing: Obtain the 512x512x3 RGB degraded image to be processed; normalize the image pixel values, converting them from the range of [0,255] to [0,1] or [-1,1].

[0079] The encoder path generates the first latent feature: The input image first passes through multiple downsampling modules of the encoder, each using the basic modules described above. Taking the first downsampling layer as an example: the input (512,512,3) passes through a 3x3 convolution with a stride of 2, and the output feature map size is (256,256,16). After several downsampling passes, the encoder outputs a high-level first latent feature at the bottom layer, which has a small spatial size and rich semantic information, with dimensions of (64,64,128).

[0080] The discriminant network path is used to generate a second latent feature: the downsampling part of the discriminant network shares the same structure as the encoder, using the exact same efficient base modules. Therefore, the computations of both are shared until they reach the same scale (64, 64, 128).

[0081] During the training phase, the discriminator network is followed by a global average pooling layer and a multilayer perceptron (MLP) classification head after this feature, outputting a vector with the same dimension as the number of degradation categories (e.g., 5 categories: deblurring, denoising, dehazing, dark enhancement, and super-resolution), and is trained using cross-entropy loss.

[0082] During the inference phase, the classification head is removed. The discriminative network directly outputs its last convolutional feature layer as a second latent feature, which also has dimensions (64, 64, 128). This feature encodes the degradation type information of the input image.

[0083] Degradation-aware feature modulation fusion: The first latent feature (64,64,128) output by the encoder is multiplied element-wise with the second latent feature (64,64,128) output by the discriminator network to obtain the preliminary modulation features.

[0084] Subsequently, the modulated features and the original first latent features are weighted and summed using a set of learnable weight parameters to generate the final fused features (64, 64, 128). This process enables the recovery network to dynamically focus on the content most relevant to the current degradation.

[0085] Image Reconstruction: The decoder consists of multiple upsampling modules, its core being a combination of an efficient base module and bilinear interpolation upsampling. The fused feature (64, 64, 128) first enters a decoder module, where convolution compresses the number of channels to 64, resulting in a feature of (64, 64, 64). Subsequently, bilinear interpolation is used to upsample this feature to (128, 128, 64). This upsampled feature is then fused with the encoder's feature at the corresponding scale (i.e., the (128, 128, 64) feature generated during downsampling). This process is repeated until the feature map is finally upsampled back to the original image size (512, 512, 16).

[0086] Finally, a 3x3 convolution is used to compress the number of channels from 16 to 3, outputting the restored image (512, 512, 3).

[0087] like Figure 6-8 As shown, the method in this embodiment has a significant enhancement effect on mixed degraded images (such as low light + blur + noise). The deblurring operation can restore contours, the low light enhancement can improve overall brightness and contrast, and the denoising operation can effectively suppress noise, with no obvious conflicts between the tasks. On the Snapdragon 8gen1 mobile platform, due to its extremely lightweight design, this embodiment processes a 256x256 input image, and the entire inference process takes an average of only 10 milliseconds, fully meeting the requirements of real-time video processing.

[0088] Example 2

[0089] In this embodiment, an electronic device is provided, including a processor and a memory, wherein the memory stores a computer program, and the processor is configured to execute the computer program to implement the real-time all-in-one image enhancement method disclosed in Embodiment 1.

[0090] Example 3

[0091] In this embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, it implements a real-time all-in-one image enhancement method as disclosed in Embodiment 1.

[0092] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A real-time all-in-one image enhancement method, characterized in that, The image enhancement method includes, Obtain the degraded image to be processed; An encoder is used to downsample the degraded image to generate a first latent feature; A discriminative network is used to process the degraded image to generate a second latent feature that characterizes the degradation type information of the degraded image; The first latent feature is modulated based on the second latent feature to generate a fused feature that incorporates degradation type information; A decoder is used to upsample the fused features to generate the restored image.

2. The real-time multi-image enhancement method according to claim 1, characterized in that, The modulation of the first latent feature based on the second latent feature includes: The first latent feature and the second latent feature are multiplied element-wise to obtain the fused feature.

3. The real-time multi-image enhancement method according to claim 1, characterized in that, The modulation of the first latent feature based on the second latent feature further includes: The features obtained through the element-wise multiplication operation are then combined with the first latent features using a weighted combination.

4. The real-time multi-image enhancement method according to claim 1, characterized in that, The encoder and the decoder constitute a U-Net framework, and The decoder is configured to, during upsampling, aggregate the intermediate features with the features generated by the encoder during downsampling, which correspond to the size of the intermediate features.

5. The real-time multi-image enhancement method according to claim 1, characterized in that, The downsampling structure of the discriminant network is constructed using the same modules as the downsampling structure of the encoder.

6. The real-time multi-image enhancement method according to any one of claims 1 to 5, characterized in that, The downsampling process in the encoder and / or the discriminant network includes at least a basic module employing the following steps: The input features are processed through the first convolutional layer; The output of the first convolutional layer is split along the channel dimension to obtain the first branch features and the second branch features; Channel attention weighting is applied to the features of the first branch; The first branch features and the second branch features, which are weighted by channel attention, are aggregated to generate aggregated features; as well as The aggregated features are subjected to a block multiplication operation to achieve feature filtering.

7. The real-time multi-image enhancement method according to claim 1, characterized in that, During the training phase, the discriminative network includes a classification head and is trained using a classification loss function to enable it to acquire degenerate recognition capabilities; during the inference phase, the classification head is removed.

8. An electronic device comprising a processor and a memory, the memory storing a computer program, the processor being configured to, when executing the computer program, implement the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images