Digital photo frame image intelligent processing method and system based on deep learning

Abstract

The invention relates to the field of image processing, and discloses a digital photo frame image intelligent processing method and system based on deep learning, and the method comprises the steps: obtaining original image data, and carrying out the preprocessing of the original image data, and obtaining a preprocessed image; extracting an image feature vector by using a convolutional neural network; based on the image feature vector, analyzing complexity and classifying to obtain an image classification result; determining an adaptive parameter set according to a classification result; acquiring light data, classifying the image feature vectors based on the light data, extracting semantic feature vectors, and adjusting the weights of the semantic feature vectors to generate dynamic strategy vectors; refining the dynamic strategy vector based on the light data to obtain a refined strategy set; and adjusting the image according to the refinement strategy set, and carrying out association comparison iterative optimization to obtain a final optimized image. According to the method, the multi-dimensional feature space is constructed, and a closed-loop iterative optimization mechanism is introduced, so that self-adaptive and high-quality image display optimization is realized.

Description

Technical Field

[0001] This invention relates to the field of image processing, and in particular to a method and system for intelligent image processing of digital photo frames based on deep learning. Background Technology

[0002] Currently, digital photo frames serve as an important tool for modern home decor and personalized display, and their image processing technology directly impacts users' visual experience and usability. With the rapid development of intelligent image processing technology, digital photo frames can automatically optimize their images through content analysis to meet diverse aesthetic needs and the requirements of different application scenarios. Their core technologies involve multiple technical fields such as image enhancement and image recognition.

[0003] In one existing technology, image processing methods for digital photo frames mainly rely on single-dimensional feature analysis, such as adjusting only the color or brightness of the image. While these methods improve the visual effect of the image to some extent, they often fail to fully capture the diverse characteristics of complex images, including texture details, semantic content, and stylistic features.

[0004] This leads to the conclusion that image processing methods with a single dimension or fixed strategy are insufficient to meet the needs of multiple scenarios and devices simultaneously. In a typical existing technology implementation, the feature analysis dimension is relatively singular, making it difficult to effectively integrate multi-dimensional features and dynamically map them to the optimal processing strategy. This results in image distortion or style inconsistency after processing, ultimately leading to poor display effects. Summary of the Invention

[0005] This invention provides a method and system for intelligent image processing of digital photo frames based on deep learning, in order to solve the problem that existing technologies rely on single-dimensional feature analysis, which leads to the inability to achieve adaptive image optimization in complex scenes and thus poor display effects.

[0006] In a first aspect, to address the aforementioned technical problems, this invention provides a deep learning-based intelligent image processing method for digital photo frames, comprising: The original image data is acquired and preprocessed to obtain the preprocessed image; Image feature vectors are extracted from the preprocessed image using a convolutional neural network; Based on the image feature vectors, the image complexity is analyzed and the images are classified to obtain the image classification results; Based on the image classification results, obtain the display requirement data of the current digital photo frame, and determine the set of adaptive parameters that match the image feature vector; Acquire light data, classify the image feature vector based on the light data, extract semantic feature vector based on the classification result, calculate the matching degree between the classification result and the semantic feature vector, adjust the weight of the semantic feature vector, and generate an optimized dynamic strategy vector. The dynamic strategy vector is refined based on the light data to obtain a refinement strategy set; the preprocessed image is adjusted according to the refinement strategy set to obtain an adjusted image; The adjusted image is correlated and compared with the image feature vector. Based on the correlation and comparison results, the adjusted image is optimized to obtain the final optimized image.

[0007] In one optional implementation, the step of acquiring the original image data and performing preprocessing to obtain a preprocessed image includes: Acquire the original image data; perform grayscale conversion and noise reduction on the original image data to obtain the first image; Based on the first image, edge detection and image segmentation operations are performed to obtain a preprocessed image divided into multiple regions.

[0008] In one optional implementation, extracting image feature vectors from the preprocessed image using a convolutional neural network includes: When the number of regions determined by the preprocessed image exceeds a preset region number threshold, feature extraction is performed on each region through a convolutional neural network to obtain feature vectors of color distribution, texture details and semantic content, resulting in a multi-dimensional feature set. Principal component analysis is used to reduce the dimensionality of the multidimensional feature set to obtain the image feature vector.

[0009] In one optional implementation, the step of analyzing image complexity based on the image feature vector and classifying the image to obtain an image classification result includes: The statistical characteristics of the texture detail components in the image feature vector are extracted using the gray-level co-occurrence matrix, and the entropy value is calculated to obtain the first complexity; When the first complexity exceeds a preset complexity threshold, K-means clustering is used to group the image feature vectors to obtain a grouped feature set; Calculate the ratio of the internal variance to the external variance of each feature cluster in the grouped feature set to obtain a variance ratio set; When at least one ratio in the variance ratio set exceeds a preset variance ratio threshold, the boundary complexity is obtained through boundary point analysis to determine the image classification result.

[0010] In one optional implementation, the step of obtaining the display requirement data of the current digital photo frame based on the image classification result and determining the set of adaptation parameters that matches the image feature vector includes: Based on the image classification results, the brightness distribution features are extracted using the grayscale histogram method, and the brightness mean and variance are calculated to obtain a brightness distribution description. When the variance in the brightness distribution description exceeds a preset brightness variance threshold, the image is converted from the RGB color space to the CIELAB color space to obtain the color gamut range parameter. The image feature vectors are grouped using K-means clustering, and the mean pixel density of each group is calculated to determine a set of display parameters that match the resolution. Based on the color gamut range parameter and the display parameter set, adaptive mapping is used to adjust the contrast and color saturation of the image to obtain the adaptive parameter set.

[0011] In one optional implementation, the steps of acquiring light data, classifying the image feature vector based on the light data, extracting semantic feature vectors based on the classification results, calculating the matching degree between the classification results and the semantic feature vectors, adjusting the weights of the semantic feature vectors, and generating an optimized dynamic strategy vector include: Light intensity data is acquired from an ambient light sensor, and the light intensity data is smoothed using a mean filtering method to obtain a light feature vector; When the mean of the light feature vector exceeds a preset light intensity threshold, the image feature vector is classified by a support vector machine to obtain a first classification result; Extract the semantic feature vector corresponding to the first classification result from the pre-established semantic feature library, and calculate the matching degree between the semantic feature vector and the first classification result; When the matching degree is lower than the preset matching degree threshold, the gradient descent method is used to adjust the weights of the semantic feature vector to generate an optimized dynamic strategy vector.

[0012] In one optional implementation, the refinement of the dynamic policy vector based on the ray data to obtain a refined policy set includes: Select a subset related to scene adaptation from the dynamic policy vector; The subsets are clustered and grouped to obtain a grouped set; Acquire texture capture information, and based on the group set, fuse the texture capture information with the light data to generate a fused feature vector; The fused feature vectors are classified using a decision tree to generate a refined strategy set.

[0013] In one optional implementation, adjusting the preprocessed image according to the thinning strategy set to obtain the adjusted image includes: Color saturation and brightness features are obtained from the preprocessed image and normalized to obtain a normalized image. Based on the normalized image, edge features and texture features are extracted and fused to obtain a fused feature set; When the average color saturation of the fused feature set is lower than a preset saturation threshold, adaptive histogram enhancement is used to adjust the color saturation. The brightness adjustment parameters are determined based on the refinement strategy set; the adjusted image is generated by combining the brightness adjustment parameters with the adjusted color saturation.

[0014] In one optional implementation, the adjusted image is correlated and compared with the image feature vector. Based on the correlation and comparison results, the adjusted image is optimized to obtain a final optimized image, including: Calculate the comparison value between the adjusted image and the image feature vector; When the comparison value is lower than the preset comparison threshold, a first comparison result is generated; When the comparison value is not lower than the preset comparison threshold, a second comparison result is generated; When the comparison result is the first comparison result, the dynamic strategy vector is iteratively updated, and the step of refining the dynamic strategy vector based on the light data is returned; When the comparison result is the second comparison result, the adjusted image is output as the final optimized image.

[0015] Secondly, the present invention provides a deep learning-based intelligent image processing system for digital photo frames, comprising: The preprocessing module is used to acquire the original image data and perform preprocessing to obtain a preprocessed image; the feature extraction module is used to extract image feature vectors from the preprocessed image using a convolutional neural network. The classification module is used to analyze the image complexity based on the image feature vector and classify the image to obtain the image classification result; The parameter determination module is used to obtain the display requirement data of the current digital photo frame based on the image classification result, and determine the set of adaptive parameters that match the image feature vector. The dynamic strategy generation module is used to acquire light data, classify the image feature vector based on the light data, extract semantic feature vectors based on the classification results, calculate the matching degree between the classification results and the semantic feature vectors, adjust the weights of the semantic feature vectors, and generate an optimized dynamic strategy vector. The strategy refinement module is used to refine the dynamic strategy vector based on the light data to obtain a refined strategy set; An image adjustment module is used to adjust the preprocessed image according to the set of thinning strategies to obtain an adjusted image; The final output module is used to compare and correlate the adjusted image with the image feature vector, and optimize the adjusted image based on the comparison results to obtain the final optimized image.

[0016] Thirdly, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the deep learning-based intelligent image processing method for digital photo frames described in any one of the above.

[0017] Fourthly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the deep learning-based intelligent image processing method for digital photo frames described above.

[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention constructs a multi-dimensional feature space that includes color distribution, texture details, and semantic content, and employs a classification method that combines gray-level co-occurrence matrix entropy analysis, K-means clustering, and variance ratio calculation to achieve multi-dimensional quantitative analysis of image complexity. The technical derivation of this method lies in the fact that existing technologies only analyze a single feature, resulting in the inability to distinguish image content and the adoption of the same optimization strategy. This invention, through multi-dimensional feature fusion and multi-level concatenated classification, can accurately quantify image complexity, thereby solving the problem that single-dimensional feature analysis in existing technologies is insufficient for representing complex images, and providing a more comprehensive feature foundation for subsequent image optimization.

[0019] (2) This invention integrates ambient light sensor data and uses a support vector machine to establish a mapping relationship between image features, ambient light, and display parameters, thereby enabling adaptive adjustment of display parameters based on image content and usage environment. The technical derivation of this method lies in the fact that, since existing methods lack a linkage mechanism between environmental perception and content understanding, this invention introduces a support vector machine as a dynamic mapping engine, enabling the system to make intelligent decisions based on "what it sees" and "what kind of lighting it is under," thus automatically adjusting the display strategy according to changes in ambient light conditions, ensuring that the digital photo frame maintains a good display effect under different lighting environments.

[0020] (3) This invention establishes a correlation and comparison mechanism between the output image and the original feature vector, and iteratively adjusts the processing strategy based on the comparison results, forming a closed-loop optimization system. The technical derivation of this method lies in the fact that traditional processing is open-loop and cannot verify the processing results. This invention, by evaluating the output quality in real time and driving iterative updates of the strategy, enables the system to have self-correction capabilities. It can approach the optimal solution through multiple iterations, thereby enabling the evaluation and feedback of the processed image quality. Through multiple iterations, the output image reaches the preset quality standards in terms of composition, detail, and color, improving the stability and reliability of image processing. Attached Figure Description

[0021] Figure 1 This is a schematic flowchart of a deep learning-based intelligent image processing method for digital photo frames provided in an embodiment of the present invention. Figure 2 This is a flowchart of the core algorithm in the intelligent image processing method for digital photo frames based on deep learning provided in the first embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of the digital photo frame image intelligent processing system based on deep learning provided in the second embodiment of the present invention. Detailed Implementation

[0022] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Reference Figure 1 The first embodiment of the present invention provides a method for intelligent image processing of digital photo frames based on deep learning, comprising the following steps: S11, acquire the original image data and perform preprocessing to obtain the preprocessed image; S12, using a convolutional neural network to extract image feature vectors from the preprocessed image; S13, Analyze the image complexity based on the image feature vector and classify the image to obtain the image classification result; S14. Based on the image classification results, obtain the display requirement data of the current digital photo frame and determine the set of adaptation parameters that match the image feature vector. S15, acquire light data, classify the image feature vector based on the light data, extract semantic feature vector based on the classification result, calculate the matching degree between the classification result and the semantic feature vector, adjust the weight of the semantic feature vector, and generate an optimized dynamic strategy vector. S16, Based on the light data, the dynamic strategy vector is refined to obtain a refined strategy set; S17, Adjust the preprocessed image according to the refinement strategy set to obtain the adjusted image; S18, perform correlation comparison between the adjusted image and the image feature vector, and optimize the adjusted image based on the correlation comparison result to obtain the final optimized image.

[0024] In step S11, the original image data is acquired and preprocessed to obtain a preprocessed image, including: Obtain the raw image data; The original image data is subjected to grayscale conversion and noise reduction processing to obtain the first image; Based on the first image, edge detection and image segmentation operations are performed to obtain a preprocessed image divided into multiple regions.

[0025] First, the original image data is converted to grayscale and denoised to obtain the first image. Grayscale conversion uses a weighted average method with the following weighting coefficients: red component 0.299, green component 0.587, and blue component 0.114. Denoising is achieved using a Gaussian blur filter, for example, a 3×3 Gaussian kernel with a standard deviation of 1.5.

[0026] Subsequently, edge detection and image segmentation were performed on the first image. Edge detection was implemented using the Canny edge detection algorithm, with a low threshold set to 50 and a high threshold set to 150. This dual-threshold setting is based on the principle that the low threshold is used to detect obvious strong edges, while the high threshold is used to identify true edge pixels; pixels located between the two are only retained if they are connected to strong edges. This parameter combination has been experimentally verified to achieve a good balance between effectively capturing the main contours of the image and suppressing noise interference.

[0027] Image segmentation employs the GrabCut algorithm based on graph cuts. This algorithm interactively optimizes the segmentation of foreground and background in an image through an iterative energy minimization process. Its core process begins with initial foreground and background region estimates provided by the user or an automated detection program, which the algorithm uses as initial conditions. Next, the algorithm models the image as a graph structure, where pixels are nodes, and the similarity between pixels and the matching degree between pixels and the region model are used as edge weights. The region model typically employs a Gaussian mixture model. Finally, the graph is segmented using a maximum flow minimum cut algorithm to find the segmentation method that minimizes the overall energy, thus outputting a refined binary segmentation mask. Ultimately, the first image is divided into multiple relatively uniform connected regions in terms of color and texture.

[0028] It's important to note that the GrabCut algorithm is a low-level segmentation algorithm based on color and texture, and its output consists of image regions lacking semantic information. The sky, trees, and grass examples listed here are merely illustrative, referring to the regions that might be visually interpreted as these semantic objects after segmentation, rather than the algorithm directly outputting such semantic labels. The segmented image is used as a preprocessed image for subsequent deep feature extraction.

[0029] In step S12, image feature vectors are extracted from the preprocessed image using a convolutional neural network, including: When the number of regions determined by the preprocessed image exceeds a preset region number threshold, feature extraction is performed on each region through a convolutional neural network to obtain feature vectors of color distribution, texture details and semantic content, resulting in a multi-dimensional feature set. Principal component analysis is used to reduce the dimensionality of the multidimensional feature set to obtain the image feature vector.

[0030] When the number of regions in the preprocessed image exceeds a preset region threshold, a convolutional neural network extracts features from each region to obtain feature vectors for color distribution, texture details, and semantic content. For example, the preset region threshold can be set to 10. This value is based on statistical analysis of a large number of digital photo frame images and has been validated on a dataset of 1000 images using the Calinski-Harabasz index. It maximizes inter-class separation, effectively distinguishes between simple and complex compositions, and avoids overprocessing simple images. Those skilled in the art can adjust this threshold within the range of 5 to 15 according to different requirements for processing speed and accuracy. For example, for the sky region, the convolutional neural network can extract the color distribution of blue tones, smooth texture, and semantic labels of the sky, generating a 128-dimensional feature vector.

[0031] Subsequently, principal component analysis (PCA) is used to reduce the dimensionality of the multidimensional feature set, reducing the 128-dimensional vector to 32 dimensions while retaining over 90% of the variance, ultimately yielding the image feature vector. Exemplarily, the specific process of PCA dimensionality reduction is as follows: First, the original 128-dimensional feature vector, composed of color distribution, texture details, and semantic content features, is Z-score standardized to eliminate the influence of differences in feature dimensions and numerical ranges. Next, the covariance matrix of the standardized feature vector is calculated, reflecting the linear correlation between different feature dimensions. Then, the eigenvalues ​​and corresponding eigenvectors of the covariance matrix are solved. The magnitude of the eigenvalue represents the magnitude of the data variance in the direction of its corresponding eigenvector. The eigenvalues ​​are sorted from largest to smallest, and their corresponding eigenvectors represent the directions with the largest data variance, i.e., the principal components. Then, the first k principal components whose cumulative variance contribution rate first reaches or exceeds 90% are selected; in this embodiment, k=32. This threshold aims to significantly reduce data dimensionality and subsequent computational complexity while retaining the vast majority of information in the original features. Finally, the original 128-dimensional feature vector is projected into a new feature space spanned by the first 32 principal components, thus obtaining the 32-dimensional image feature vector after dimensionality reduction.

[0032] The convolutional neural network described is a deep learning model capable of automatically learning and extracting hierarchical features from images. For example, this invention uses a pre-trained VGG-16 model as a feature extractor and fine-tunes it to adapt to the characteristics of digital photo frame images. This network structure includes 13 convolutional layers, 5 pooling layers, and 3 fully connected layers. The convolutional layers all use small 3×3 convolutional kernels and employ the ReLU function as the activation function to introduce non-linear transformation. The pooling layers all use 2×2 window max pooling to progressively reduce the feature map size while maintaining feature invariance. For the feature extraction task of this invention, the final classification layer of the original model, i.e., the third fully connected layer, is removed. The output of the second fully connected layer is used as the deep feature representation of each image region, and then a dedicated feature transformation layer converts this high-dimensional feature into a 128-dimensional feature vector. The model fine-tuning process is supervised learning, using a self-collected dataset of digital photo frame images containing various scenes such as portraits, landscapes, and still lifes, supplemented with image category or quality rating labels. During training, preprocessed images are used as input, and image category or quality score is used as the supervision objective. Cross-entropy loss or mean squared error is used as the loss function, and stochastic gradient descent is employed. The initial learning rate is set to 0.001, and the momentum factor is 0.9. Multiple iterations are performed on the training data, updating the network weights through backpropagation, until the model's loss function converges and its performance is stable on the validation set. The trained model can automatically extract highly abstract image feature vectors from the input preprocessed images, suitable for subsequent complexity analysis and policy generation. Those skilled in the art will understand that other modern network architectures such as ResNet or Inception can be used to replace the VGG-16 model.

[0033] In step S13, the image complexity is analyzed based on the image feature vector, and the image is classified to obtain the image classification result, including: The statistical characteristics of the texture detail components in the image feature vector are extracted using the gray-level co-occurrence matrix, and the entropy value is calculated to obtain the first complexity; When the first complexity exceeds a preset complexity threshold, K-means clustering is used to group the image feature vectors to obtain a grouped feature set; Calculate the ratio of the internal variance to the external variance of each feature cluster in the grouped feature set to obtain a variance ratio set; When at least one ratio in the variance ratio set exceeds a preset variance ratio threshold, the boundary complexity is obtained through boundary point analysis to determine the image classification result.

[0034] First, the statistical characteristics of texture detail components in the image feature vector are extracted using a gray-level co-occurrence matrix (GLCM), and the entropy value is calculated to obtain the first complexity. Exemplarily, this invention uses a GLCM to quantify the statistical characteristics of the texture detail components. The construction of this matrix is ​​based on two core parameters: the calculation direction and the pixel distance. The calculation direction is used to capture texture features at different angles; typically, 0°, 45°, 90°, and 135° are selected for calculation, and the average of the results from the four directions is taken as the rotation-invariant texture descriptor. The pixel distance is used to define the observation scale of the texture, typically set to 1 pixel to capture microscopic texture structures; for images containing more macroscopic textures, the distance parameter can be adjusted to 3 or 5 pixels as appropriate. The establishment of these parameter combinations is based on common knowledge in the field of digital image texture analysis and is verified through testing on standard texture libraries to ensure its discriminative power for various texture patterns. After generating the GLCM, its entropy value is calculated to quantify the disorder and information content of the texture; a higher entropy value indicates a more complex texture.

[0035] Subsequently, for example, the preset complexity threshold is set to 5.0. This threshold is determined by analyzing the correlation between texture entropy values ​​and manually labeled complexity tags on an image dataset. When the entropy value is higher than 5.0, the image has a greater than 90% probability of being manually judged as complex; therefore, the complexity threshold is set to 5.0. When the first complexity exceeds the preset complexity threshold of 5.0, the image feature vectors are grouped using the K-means clustering algorithm to obtain grouped feature sets.

[0036] For example, the specific implementation process of the K-means clustering algorithm is as follows: First, the number of clusters K needs to be determined. This invention uses the elbow method to determine the K value. By calculating the sum of squares within the clusters corresponding to different K values ​​from 2 to 10, a curve showing the relationship between the K value and the sum of squares within the clusters is plotted. The K value corresponding to the inflection point of the curve, i.e., the elbow position, is selected as the optimal number of clusters to achieve a balance between clustering effect and model complexity. Second, during the clustering process, Euclidean distance is used as a metric to quantify the similarity of feature vectors in space. Vectors that are closer in distance are considered more likely to belong to the same cluster. For the 32-dimensional feature vectors after PCA dimensionality reduction, Z-score standardization is performed again before clustering to eliminate the influence of variance differences in each principal component dimension and ensure the objectivity of the clustering results. At the start of clustering, K cluster centers are randomly initialized. Subsequently, the following two steps are iteratively executed until convergence: First, each feature vector is assigned to the cluster to which the nearest cluster center belongs; second, the mean of all feature vectors in each cluster is recalculated, and this mean is used as the new cluster center. The termination condition for the iteration is set as follows: the movement distance of the cluster center point is less than a preset threshold of 10⁻. 4The iteration count may be increased to the maximum of 100. The resulting stable set of grouped features is then used for subsequent variance ratio calculations.

[0037] Then, the ratio of the internal variance to the external variance of each feature cluster in the grouped feature set is calculated to obtain the variance ratio set.

[0038] For example, the preset variance ratio threshold is set to 10.0. This threshold is determined based on statistical analysis of a large number of sample images: the variance ratio after clustering is calculated and compared with manually labeled "complex image." When the variance ratio is higher than 10.0, the accuracy rate of classifying an image as a complex image exceeds 85%, thus this value is selected as the distinguishing criterion. This threshold works in conjunction with the entropy threshold to form a cascaded judgment, further filtering out images with high content dispersion and distinct features. When at least one ratio in the variance ratio set exceeds this threshold, boundary complexity analysis is triggered.

[0039] For example, the boundary complexity is obtained by analyzing the degree of disorder in the distribution of feature cluster boundary regions. Specifically, firstly, the two feature clusters that are closest to each other are identified from all feature clusters. Then, a set of boundary points is determined in each of these two feature clusters. The criteria for determining a boundary point is that the distance from the point to the center of the adjacent cluster is less than or equal to 1.5 times its distance to the center of its own cluster. Next, the ratio of the average distance of these boundary points relative to their own cluster center to their distance to the center of the adjacent cluster is calculated. The higher the ratio, the more blurred and ambiguous the boundary points are, and the higher the boundary complexity. For example, in a complex landscape image containing sky and trees, the smooth blue feature vector of the sky cluster and the green texture feature vector of the tree cluster will be mixed together in the boundary region. The calculated ratio will be significantly higher than that of a simple image of a single object against a solid-color background, thus revealing a higher boundary complexity. The preset threshold for boundary complexity is set to 0.6. This threshold is determined by analyzing the statistical relationship between the boundary complexity ratio and manually labeled image complexity on an image dataset. When the boundary complexity is higher than 0.6, the image has more than 85% probability of being classified as a complex image. Finally, when the boundary complexity exceeds its preset threshold, the image classification result is determined to be a complex image.

[0040] In step S14, based on the image classification result, the display requirement data of the current digital photo frame is obtained, and an adaptation parameter set matching the image feature vector is determined, including: Based on the image classification results, the brightness distribution features are extracted using grayscale histograms, and the brightness mean and variance are calculated to obtain a brightness distribution description. When the variance in the brightness distribution description exceeds a preset brightness variance threshold, the image is converted from the RGB color space to the CIELAB color space to obtain the color gamut range parameter. The image feature vectors are grouped using K-means clustering, and the mean pixel density of each group is calculated to determine a set of display parameters that match the resolution. Based on the color gamut range parameter and the display parameter set, adaptive mapping is used to adjust the contrast and color saturation of the image to obtain the adaptive parameter set.

[0041] First, based on the image classification results, the brightness distribution features of the image are extracted using a grayscale histogram, and the mean and variance of the brightness are calculated to obtain a brightness distribution description. The introduction of the image classification results aims to provide prior guidance for subsequent parameter determination. Specifically, when an image is classified as a complex type, the system will initiate a more refined brightness analysis and color space conversion process to ensure adaptive optimization for complex content; while for simple types of images, the processing flow can be appropriately simplified to improve efficiency.

[0042] For example, when the variance described by the luminance distribution exceeds a preset luminance variance threshold, the preset luminance variance threshold is an empirical value set by analyzing the statistical characteristics of luminance distributions in a large standard image library, typically set to 20, to effectively distinguish between images with a flat dynamic range and those with a wide dynamic range. The image is converted from the RGB color space to the CIELAB color space using a standard conversion formula established by the International Commission on Illumination. The specific conversion process is as follows: first, the linear RGB values ​​are converted to the XYZ color space using the following formula: Where R, G, and B represent the normalized red, green, and blue component values, respectively, and X, Y, and Z are the transformed tristimulus values. Subsequently, a nonlinear transformation is used to obtain the L, a, and b components in the CIELAB color space, where L represents lightness, a represents red-green hue, and b represents yellow-blue hue. The color gamut parameter is obtained by calculating the color distribution range of the image on the ab plane of the CIELAB color space, specifically the maximum and minimum values ​​of the a and b components, used to quantify the color richness of the image.

[0043] For example, image feature vectors are grouped using the K-means clustering algorithm. The K-means clustering process is as described in step S13 above, where the K value is set between 3 and 5 based on the complexity of the image content. The mean pixel density of each group is calculated to determine a resolution parameter set that matches the resolution. The resolution parameter set is determined by mapping the mean pixel density to a preset display resolution level. Specifically, the mapping process is as follows: when the mean pixel density is in the range of 0-0.3, it corresponds to the lowest display sharpness parameter level 1; when it is in the range of 0.3-0.6, it corresponds to the medium display sharpness parameter level 2; and when it is higher than 0.6, it corresponds to the highest display sharpness parameter level 3. For example, for landscape images containing rich details, the mean pixel density is usually higher than 0.6, thus corresponding to a display sharpness parameter level of 3; while for portrait images with a simple background, the mean pixel density is usually lower than 0.3, corresponding to a display sharpness parameter level of 1.

[0044] For example, based on the color gamut range parameter and resolution parameter set, the contrast and color saturation of the image are adjusted using a predefined multidimensional lookup table to obtain an adaptation parameter set. The predefined multidimensional lookup table is constructed as follows: During the development phase, a large number of sample images containing different color gamut and resolution characteristics are collected. Professional image designers manually set the optimal contrast and saturation adjustment amounts for each sample, thereby establishing a mapping relationship from color gamut range and resolution parameters to contrast and saturation adjustment amounts. Specific rules include: when the color gamut range parameter shows a richly colored image and the resolution parameter is level 3, the lookup table outputs a contrast adjustment coefficient of 1.2 and a saturation adjustment coefficient of 1.1; when the color gamut range parameter shows a monotonous color and the resolution parameter is level 1, the lookup table outputs a contrast adjustment coefficient of 1.0 and a saturation adjustment coefficient of 1.0. For example, for a brightly colored, highly detailed, high-saturation landscape image, the system will obtain the improved contrast and saturation parameters through the lookup table, thereby enhancing the visual impact of the image; while for a monotonous, simple document image, the system will maintain the original contrast and saturation settings to avoid over-processing and introducing distortion. The set of adaptation parameters includes the specific contrast adjustment coefficients and saturation adjustment coefficients generated accordingly.

[0045] In step S15, light data is acquired, the image feature vector is classified based on the light data, a semantic feature vector is extracted based on the classification result, the matching degree between the classification result and the semantic feature vector is calculated, the weights of the semantic feature vector are adjusted, and an optimized dynamic strategy vector is generated, including: Light intensity data is acquired from an ambient light sensor, and the light intensity data is smoothed using a mean filtering method to obtain a light feature vector; When the mean of the light feature vector exceeds a preset light intensity threshold, the image feature vector is classified by a support vector machine to obtain a first classification result; Extract the semantic feature vector corresponding to the first classification result from the pre-established semantic feature library, and calculate the matching degree between the semantic feature vector and the first classification result; When the matching degree is lower than the preset matching degree threshold, the gradient descent method is used to adjust the weights of the semantic feature vector to generate an optimized dynamic strategy vector.

[0046] For example, a light intensity data sequence, in lux, is acquired from an ambient light sensor. The light intensity data is smoothed using a mean filtering method to eliminate instantaneous fluctuations. The mean filtering window size is set to 5 sampling points. The light feature vector is constructed as follows: first, the arithmetic mean of the light intensity data within the window is calculated as the smoothed light intensity value; then, the smoothed values ​​from 10 consecutive time series are arranged chronologically to form a 10-dimensional feature vector, representing the stable illumination state over a period of time.

[0047] For example, the preset light intensity threshold is set to 100 lux. This threshold corresponds to the typical indoor bright ambient light level and is used to trigger a specific optimization strategy for high-brightness environments. When the mean of the light feature vector exceeds this threshold, the image feature vector is classified using a support vector machine model to obtain a first classification result, such as a high-brightness environment.

[0048] For example, a standard semantic feature vector corresponding to the first classification result is extracted from a pre-established semantic feature library. The process of determining the ideal parameter vector is as follows: based on the first classification result, the pre-stored standard display parameter vector corresponding to the environment type is directly retrieved from the semantic feature library. For example, for the high-brightness environment classification result, the corresponding standard semantic feature vector is extracted, which contains a luminance component of 500 nits and a color temperature component of 6500K. The matching degree between the semantic feature vector and the first classification result is calculated. The matching degree is calculated as follows: first, the luminance component and color temperature component in the semantic feature vector are respectively subjected to minimum-maximum normalization to eliminate the influence of different dimensions, mapping the luminance component to the range of 0 to 1, and mapping the color temperature component to the range of 0 to 1. Then, the Euclidean distance between the normalized semantic feature vector and the ideal parameter vector is calculated. Finally, the Euclidean distance is converted into a percentage matching degree using a formula, where the maximum normalized distance is set to 1.5, which is an empirical value determined by analyzing the maximum possible distance between the semantic feature vector and the ideal parameter vector under all environment types.

[0049] For example, the preset matching degree threshold is set to 80%. This threshold is selected after evaluating the policy effect on a validation set containing 1000 samples. Specifically, the selection process involves testing the system's optimization effect under different candidate thresholds and finally selecting the threshold that gives the highest overall system performance score as the preset matching degree threshold. When the matching degree is lower than the preset matching degree threshold, the gradient descent method is used to adjust the weights of the semantic feature vector. The specific process of the gradient descent method is as follows: the loss function is defined as the square of the matching degree error, the learning rate is set to 0.01 during the weight update process, and the maximum number of iterations is set to 100 to minimize the matching degree error. By iteratively updating the weight coefficients of the semantic feature vector, an optimized dynamic policy vector is generated, which includes the adjusted brightness weight and color temperature weight. If the matching degree is greater than or equal to the preset matching degree threshold, the semantic feature vector is directly used as the dynamic policy vector.

[0050] For example, the Support Vector Machine (SVM) is a commonly used supervised learning classification model. The training process employs a sequential minimum optimization algorithm, with core parameters including the radial basis function kernel coefficients and the penalty factor. Those skilled in the art can train the SVM model using this standard training method by collecting a training sample set containing 100,000 image feature vectors and corresponding environment type labels, enabling it to learn the classification boundaries between image features and environment types. The training samples cover various typical lighting environments, including indoor, outdoor, strong light, and weak light.

[0051] For example, the semantic feature library is stored in the form of a database, realizing the mapping from environment type to display parameter vector. The process of building the semantic feature library is as follows: during the system development phase, data is collected under various standard lighting environments, and experts calibrate the optimal display parameters for each environment type based on visual perception experiments. Finally, the correspondence between environment type and display parameters is persistently stored in the database.

[0052] In step S16, the dynamic strategy vector is refined based on the light data to obtain a refined strategy set, including: Select a subset related to scene adaptation from the dynamic policy vector; The subsets are clustered and grouped to obtain a grouped set; Acquire texture capture information, and based on the group set, fuse the texture capture information with the light data to generate a fused feature vector; The fused feature vectors are classified using a decision tree to generate a refined strategy set.

[0053] For example, a subset related to scene adaptation is selected from the optimized dynamic strategy vector. The determination logic for the subset related to scene adaptation is as follows: parameters that are sensitive to changes in ambient lighting conditions and show significant adjustment amplitudes in historical data are selected. The specific quantitative determination rules include two conditions: first, the coefficient of variation of the parameter under different ambient lighting conditions is greater than 0.3; second, the standard deviation of the adjustment amplitude of the parameter in historical adjustment records is greater than 15% of its mean. Parameters that meet both conditions are considered related to scene adaptation, typically including global brightness, global contrast, and global color temperature.

[0054] For example, the subset is clustered and grouped using the K-means clustering method to obtain the grouped sets. The specific implementation process and parameter settings are the same as those in step S13 above. Here, based on the dimension and distribution characteristics of the subset parameters, the number of clusters K is usually set between 2 and 4 to identify a limited number of typical parameter combination patterns.

[0055] For example, texture capture information of the image is extracted, and based on the grouping set, the texture capture information and lighting data are fused, and a weighted average method is used to generate a fused feature vector. The texture capture information is specifically defined here as the global texture complexity index of the image, obtained by calculating the average entropy value of the entire image through the gray-level co-occurrence matrix. The fusion process is as follows: first, the data of three different dimensions—the policy parameter subset, texture complexity, and mean lighting intensity—are Z-score standardized to eliminate the influence of dimensions, and then the vectors are concatenated to form a unified fused feature vector. This fused feature vector contains three components: standardized policy parameters, standardized texture complexity, and standardized mean lighting intensity.

[0056] For example, a decision tree is used to classify the fused feature vectors to generate a refinement strategy set. The decision tree is constructed using the CART algorithm, with node splitting based on minimizing the Gini coefficient. The maximum tree depth is preset to 5 to control model complexity. The training process of this decision tree is as follows: using a historical dataset containing fused feature vectors and expert-annotated recommended refinement strategy labels, a tree structure is constructed through recursive binary search until the purity of leaf nodes exceeds 95% or the depth limit is reached. The quantification standard for leaf node purity is the proportion of samples in that leaf node belonging to the dominant class. Finally, each leaf node is associated with a specific strategy label. In application, the fused feature vectors, starting from the root node, are assigned to a leaf node based on the comparison of the values ​​of each dimension with the node splitting threshold. The strategy label associated with that leaf node is the generated refinement strategy set. For example, the decision tree might classify a scene as medium lighting with high texture and output a refinement strategy set that increases contrast by 8% and maintains a color temperature of 6000K.

[0057] In step S17, the preprocessed image is adjusted according to the thinning strategy set to obtain the adjusted image, including: Color saturation and brightness features are obtained from the preprocessed image and normalized to obtain a normalized image. Based on the normalized image, edge features and texture features are extracted and fused to obtain a fused feature set; When the average color saturation of the fused feature set is lower than a preset saturation threshold, adaptive histogram enhancement is used to adjust the color saturation. The brightness adjustment parameters are determined based on the refined strategy set; The adjusted image is generated by combining the brightness adjustment parameters with the adjusted color saturation.

[0058] For example, color saturation and brightness features are obtained from a preprocessed image. The brightness features are obtained by calculating the histogram distribution of the V channel of the image in the HSV color space.

[0059] For example, color saturation and brightness features are obtained from a preprocessed image. The brightness features are obtained by calculating the histogram distribution of the V channel of the image in the HSV color space.

[0060] For example, the color saturation and brightness features are normalized using a histogram equalization method to obtain a normalized image. Based on the normalized image, edge features and texture features are extracted and fused to obtain a first feature set. The edge features are obtained by calculating the image gradient magnitude and taking the mean value using the Sobel operator; the texture features are obtained by calculating the contrast of the gray-level co-occurrence matrix. The fusion process combines the edge feature values, texture feature values, and the mean color saturation value into a three-dimensional feature vector, forming the first feature set.

[0061] For example, the preset saturation threshold is set to a normalized value of 0.6. This threshold is determined based on research on human visual perception; images below this value are generally considered to have dull colors and require saturation enhancement to improve visual appeal. When the average color saturation of the first feature set is lower than the preset saturation threshold of 0.6, an adaptive histogram enhancement method is used to adjust the color saturation, resulting in a second feature set. The specific adjustment process of the adaptive histogram enhancement method is as follows: the image is converted to the HSV color space, and the saturation S channel is processed. First, the image is divided into 8x8 local regions, and a saturation histogram is calculated in each region. Then, each histogram is cropped, and the portion exceeding the preset cropping limit of 2.5 is redistributed to each histogram interval. Finally, histogram equalization is performed on each region, and bilinear interpolation is used to eliminate blockiness.

[0062] For example, a brightness adjustment parameter, such as 150 nits, is determined based on a refinement strategy set. The brightness adjustment parameter is then combined with the adjusted color saturation to adjust the corresponding display parameters of the image, generating the adjusted image. The combination process involves: in the HSV color space, linearly scaling the pixel values ​​of the V channel using the brightness adjustment parameter, while simultaneously loading the adaptively enhanced S channel values; finally, converting the image from HSV space back to RGB space to complete the adjustment of the display parameters and generate the final adjusted image.

[0063] In step S18, the adjusted image and the image feature vector are correlated and compared. Based on the correlation and comparison results, the adjusted image is optimized to obtain the final optimized image, including: Calculate the comparison value between the adjusted image and the image feature vector; When the comparison value is lower than the preset comparison threshold, a first comparison result is generated; When the comparison value is not lower than the preset comparison threshold, a second comparison result is generated; When the comparison result is the first comparison result, the dynamic strategy vector is iteratively updated, and the step of refining the dynamic strategy vector based on the light data is returned; When the comparison result is the second comparison result, the adjusted image is output as the final optimized image.

[0064] For example, a comparison value between the adjusted image and the original image feature vector is calculated. This comparison value is measured by calculating the cosine similarity between the feature vectors of the adjusted image and the original image. The feature vector acquisition process for the adjusted image is the same as that for the original image feature vector; that is, features are extracted from the adjusted image using the same convolutional neural network to obtain a feature representation with the same dimension as the original image feature vector. The closer this value is to 1, the more similar the two are in the feature space.

[0065] For example, the preset comparison threshold is set to 0.85. This threshold is a balance point selected after plotting a precision-recall curve on a validation set, which consists of sample images collected during the system development phase, covering various typical lighting conditions and image types, and undergoing quality assessment to ensure representativeness. At this threshold, a high optimization success rate and a low false positive rate can be simultaneously guaranteed. When the comparison value is below 0.85, generating a first comparison result indicates that readjustment is needed. When the comparison value is not lower than 0.85, generating a second comparison result indicates that the standard is met.

[0066] For example, when the alignment result is the first alignment result, the dynamic policy vector is iteratively updated, and the step of refining the dynamic policy vector based on the ray data is returned. The specific process of the iterative update is as follows: using gradient descent to minimize the alignment error, the parameters in the dynamic policy vector are fine-tuned, with the learning rate set to 0.05. The iteration ends when the alignment value reaches 0.85 or the number of iterations reaches a preset upper limit of 50.

[0067] For example, when the comparison result is the second comparison result, the adjusted image is output as the final optimized image.

[0068] Reference Figure 2 The first embodiment of the present invention provides a flowchart of the core algorithm for implementing the method of the present invention. The flowchart adopts a three-column parallel layout, which clearly reveals the internal working principle and data interaction relationship of the three key algorithm modules in the intelligent image processing process, and constitutes the algorithmic foundation of the technical solution of the present invention.

[0069] The left side is the complexity analysis module, which adopts a three-level serial architecture of "texture entropy -> feature clustering -> boundary analysis" to screen layer by layer and accurately identify complex images, providing a basis for subsequent differentiated processing.

[0070] The central section is the dynamic strategy generation module, which serves as the system's intelligent decision-making hub. Based on image features and ambient lighting data, it generates an adaptive strategy vector that highly matches the current scene through classification, retrieval, and online fine-tuning, achieving a leap from fixed rules to data-driven decision-making.

[0071] The right side shows the closed-loop iterative control module, which is the core component ensuring output quality. It evaluates the effect by comparing the similarity between the adjusted image and the original features. If the standard is not met, it triggers the iterative update of the strategy vector and reprocesses the image, forming an automated optimization loop of "execution-evaluation-feedback-correction" to ensure the stability and reliability of the final result.

[0072] Reference Figure 3 The second embodiment of the present invention provides a digital photo frame image intelligent processing system based on deep learning, comprising: The preprocessing module is used to acquire raw image data and perform preprocessing to obtain a preprocessed image; The feature extraction module is used to extract image feature vectors from the preprocessed image using a convolutional neural network; The classification module is used to analyze the image complexity based on the image feature vector and classify the image to obtain the image classification result; The parameter determination module is used to obtain the display requirement data of the current digital photo frame based on the image classification result, and determine the set of adaptive parameters that match the image feature vector. The dynamic strategy generation module is used to acquire light data, classify the image feature vector based on the light data, extract semantic feature vectors based on the classification results, calculate the matching degree between the classification results and the semantic feature vectors, adjust the weights of the semantic feature vectors, and generate an optimized dynamic strategy vector. The strategy refinement module is used to refine the dynamic strategy vector based on the light data to obtain a refined strategy set; An image adjustment module is used to adjust the preprocessed image according to the set of thinning strategies to obtain an adjusted image; The final output module is used to compare and correlate the adjusted image with the image feature vector, and optimize the adjusted image based on the comparison results to obtain the final optimized image.

[0073] It should be noted that the deep learning-based intelligent image processing system for digital photo frames provided in this embodiment of the invention is used to execute all the process steps of the deep learning-based intelligent image processing method for digital photo frames in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0074] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps described in the method embodiments above.

[0075] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0076] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0077] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0078] The memory can be used to store the computer programs and modules. The processor implements various functions of the electronic device by running or executing the computer programs and modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0079] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0080] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0081] 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 descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that 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 for those skilled in the art.

Claims

1. A method for intelligent image processing of digital photo frames based on deep learning, characterized in that, include: The original image data is acquired and preprocessed to obtain the preprocessed image; Image feature vectors are extracted from the preprocessed image using a convolutional neural network; Based on the image feature vectors, the image complexity is analyzed and the images are classified to obtain the image classification results; Based on the image classification results, obtain the display requirement data of the current digital photo frame, and determine the set of adaptive parameters that match the image feature vector; Acquire light data, classify the image feature vector based on the light data, extract semantic feature vector based on the classification result, calculate the matching degree between the classification result and the semantic feature vector, adjust the weight of the semantic feature vector, and generate an optimized dynamic strategy vector. The dynamic strategy vector is refined based on the light data to obtain a refined strategy set. The preprocessed image is adjusted according to the set of refinement strategies to obtain the adjusted image; The adjusted image is correlated and compared with the image feature vector. Based on the correlation and comparison results, the adjusted image is optimized to obtain the final optimized image.

2. The intelligent image processing method for digital photo frames based on deep learning according to claim 1, characterized in that, The process of acquiring raw image data and preprocessing it to obtain a preprocessed image includes: Obtain the raw image data; The original image data is subjected to grayscale conversion and noise reduction processing to obtain the first image; Based on the first image, edge detection and image segmentation operations are performed to obtain a preprocessed image divided into multiple regions.

3. The intelligent image processing method for digital photo frames based on deep learning according to claim 1, characterized in that, The step of extracting image feature vectors from the preprocessed image using a convolutional neural network includes: When the number of regions determined by the preprocessed image exceeds a preset region number threshold, feature extraction is performed on each region through a convolutional neural network to obtain feature vectors of color distribution, texture details and semantic content, resulting in a multi-dimensional feature set. Principal component analysis is used to reduce the dimensionality of the multidimensional feature set to obtain the image feature vector.

4. The intelligent image processing method for digital photo frames based on deep learning according to claim 1, characterized in that, The process of analyzing image complexity based on the image feature vector and classifying the image to obtain the image classification result includes: The statistical characteristics of the texture detail components in the image feature vector are extracted using the gray-level co-occurrence matrix, and the entropy value is calculated to obtain the first complexity; When the first complexity exceeds a preset complexity threshold, K-means clustering is used to group the image feature vectors to obtain a grouped feature set; Calculate the ratio of the internal variance to the external variance of each feature cluster in the grouped feature set to obtain a variance ratio set; When at least one ratio in the variance ratio set exceeds a preset variance ratio threshold, the boundary complexity is obtained through boundary point analysis to determine the image classification result.

5. The intelligent image processing method for digital photo frames based on deep learning according to claim 1, characterized in that, The step of obtaining the display requirement data of the current digital photo frame based on the image classification result and determining the set of adaptation parameters that matches the image feature vector includes: Based on the image classification results, the brightness distribution features are extracted using the grayscale histogram method, and the brightness mean and variance are calculated to obtain a brightness distribution description. When the variance in the brightness distribution description exceeds a preset brightness variance threshold, the image is converted from the RGB color space to the CIELAB color space to obtain the color gamut range parameter. The image feature vectors are grouped using K-means clustering, and the mean pixel density of each group is calculated to determine a set of display parameters that match the resolution. Based on the color gamut range parameter and the display parameter set, adaptive mapping is used to adjust the contrast and color saturation of the image to obtain the adaptive parameter set.

6. The intelligent image processing method for digital photo frames based on deep learning according to claim 1, characterized in that, The process of acquiring light data, classifying the image feature vector based on the light data, extracting semantic feature vectors based on the classification results, calculating the matching degree between the classification results and the semantic feature vectors, adjusting the weights of the semantic feature vectors, and generating an optimized dynamic strategy vector includes: Light intensity data is acquired from an ambient light sensor, and the light intensity data is smoothed using a mean filtering method to obtain a light feature vector; When the mean of the light feature vector exceeds a preset light intensity threshold, the image feature vector is classified by a support vector machine to obtain a first classification result; Extract the semantic feature vector corresponding to the first classification result from the pre-established semantic feature library, and calculate the matching degree between the semantic feature vector and the first classification result; When the matching degree is lower than the preset matching degree threshold, the gradient descent method is used to adjust the weights of the semantic feature vector to generate an optimized dynamic strategy vector.

7. The intelligent image processing method for digital photo frames based on deep learning according to claim 1, characterized in that, The refinement of the dynamic strategy vector based on the light data yields a refined strategy set, including: Select a subset related to scene adaptation from the dynamic policy vector; The subsets are clustered and grouped to obtain a grouped set; Acquire texture capture information, and based on the group set, fuse the texture capture information with the light data to generate a fused feature vector; The fused feature vectors are classified using a decision tree to generate a refined strategy set.

8. The intelligent image processing method for digital photo frames based on deep learning according to claim 1, characterized in that, The step of adjusting the preprocessed image according to the set of thinning strategies to obtain the adjusted image includes: Color saturation and brightness features are obtained from the preprocessed image and normalized to obtain a normalized image. Based on the normalized image, edge features and texture features are extracted and fused to obtain a fused feature set; When the average color saturation of the fused feature set is lower than a preset saturation threshold, adaptive histogram enhancement is used to adjust the color saturation. The brightness adjustment parameters are determined based on the refined strategy set; The adjusted image is generated by combining the brightness adjustment parameters with the adjusted color saturation.

9. The intelligent image processing method for digital photo frames based on deep learning according to claim 1, characterized in that, The adjusted image is correlated and compared with the image feature vector. Based on the correlation and comparison results, the adjusted image is optimized to obtain the final optimized image, including: Calculate the comparison value between the adjusted image and the image feature vector; When the comparison value is lower than the preset comparison threshold, a first comparison result is generated; When the comparison value is not lower than the preset comparison threshold, a second comparison result is generated; When the comparison result is the first comparison result, the dynamic strategy vector is iteratively updated, and the step of refining the dynamic strategy vector based on the light data is returned; When the comparison result is the second comparison result, the adjusted image is output as the final optimized image.

10. A deep learning-based intelligent image processing system for digital photo frames, characterized in that, include: The preprocessing module is used to acquire raw image data and perform preprocessing to obtain a preprocessed image; The feature extraction module is used to extract image feature vectors from the preprocessed image using a convolutional neural network; The classification module is used to analyze the image complexity based on the image feature vector and classify the image to obtain the image classification result; The parameter determination module is used to obtain the display requirement data of the current digital photo frame based on the image classification result, and determine the set of adaptive parameters that match the image feature vector. The dynamic strategy generation module is used to acquire light data, classify the image feature vector based on the light data, extract semantic feature vectors based on the classification results, calculate the matching degree between the classification results and the semantic feature vectors, adjust the weights of the semantic feature vectors, and generate an optimized dynamic strategy vector. The strategy refinement module is used to refine the dynamic strategy vector based on the light data to obtain a refined strategy set; An image adjustment module is used to adjust the preprocessed image according to the set of thinning strategies to obtain an adjusted image; The final output module is used to compare and correlate the adjusted image with the image feature vector, and optimize the adjusted image based on the comparison results to obtain the final optimized image.

Images