An adaptive digital media content intelligent cropping and composition optimization method based on deep learning

Abstract

This invention discloses an adaptive intelligent cropping and composition optimization method for digital media content based on deep learning. The steps are as follows: S1, short video frames and live streaming frames are acquired through a media material acquisition and pattern recognition module; S2, a dual-branch deep feature extraction module is used with a dual-path parallel structure; S3, the output features of the dual-branch deep feature extraction module are associated with the output features of the dual-branch deep feature extraction module through a dynamic weight fusion module; S4, a preset ratio adaptation structure and an optimal solution structure are called through an adaptive cropping solution module; S5, a composition optimization and output module is used with a multi-level serial unit structure. This solution, by leveraging the cooperation between the dual-branch deep feature extraction module and the dynamic weight fusion module, can simultaneously acquire key content information and aesthetic layout information of the image, overcoming the technical limitations of traditional cropping structures in achieving a balance between content retention and visual aesthetics.

Description

Technical Field

[0001] This invention relates to the technical field of an adaptive digital media content intelligent cropping and composition optimization method based on deep learning, specifically an adaptive digital media content intelligent cropping and composition optimization method based on deep learning. Background Technology

[0002] With the continued rapid development of new media industries such as short videos and live streaming, the screen specifications and display ratios of terminal devices such as mobile phones, tablets, and smart screens are becoming increasingly diverse. Media materials need to undergo image cropping before being uploaded and distributed, making intelligent automatic cropping a necessary step in the industrial production of digital content. Currently, deep learning technology is being widely introduced into the field of intelligent image cropping, and various automatic cropping solutions based on neural networks are being implemented and applied.

[0003] Most of the currently available intelligent image cropping technologies follow the basic logic of traditional center-cropping. The cropping judgment relies solely on the salience information of a single image to complete the boundary screening, and the overall technical architecture has many shortcomings.

[0004] First, existing solutions lack aesthetic composition criteria. During cropping, they prioritize geometric centering of the image, often removing key content such as portraits, subtitles, and brand logos. Intentionally preserving the main content can easily lead to an unbalanced image and jarring composition, failing to balance content integrity and visual appeal. Second, conventional cropping systems employ a single deployment architecture, either only compatible with large-scale offline processing on cloud servers or capable of lightweight computation on terminal devices. They cannot simultaneously handle batch material processing and real-time live video cropping, requiring companies to purchase two separate processing devices, resulting in high hardware and maintenance costs. Third, for continuous frame dynamic materials such as live streams and short videos, existing cropping algorithms independently calculate the cropping position for each frame, lacking an inter-frame coordinate constraint mechanism. This leads to significant deviations in cropping areas between adjacent frames, causing frequent flickering and shaking during video playback. Fourth, existing cropping processes stop at image size cropping. After cropping, there are no standardized image restoration steps. Imperfections such as edge defects and image center of gravity shifts can only be corrected manually in post-production, significantly increasing the manual labor time required for content creation.

[0005] In summary, existing intelligent cropping technologies for digital media are ill-suited to the diverse application scenarios of today's new media. They exhibit significant technical deficiencies in areas such as critical content protection, cross-deployment compatibility, dynamic video stability, and the regularity of finished product images. Therefore, there is an urgent need to design a new intelligent cropping optimization solution. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a deep learning-based adaptive intelligent cropping and composition optimization method for digital media content, solving the problems mentioned in the background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based adaptive intelligent cropping and composition optimization method for digital media content, the steps of which are as follows: S1. Collect original media materials such as short video frames, live broadcast images, and static images through the media material acquisition and pattern recognition module, identify the current device operating environment, and adaptively switch between cloud batch processing mode and device real-time processing mode according to the differences in device environment. S2. Through the dual-branch deep feature extraction module, a dual-path parallel structure is adopted to simultaneously and independently extract key content features and aesthetic composition features. The key content features are used to locate the core target area that needs to be retained in the picture, and the aesthetic composition features are used to characterize the visual balance and layout of the picture. S3. By associating the output features of the dual-branch deep feature extraction module with the dynamic weight fusion module, the weight ratio of the two types of features is matched according to the differences in the content type of the image, and a unique cropping judgment standard adapted to the current image is generated. S4. The adaptive cropping solution module calls the preset ratio adaptation structure and the optimal solution structure to adapt to any custom screen display ratio and locks the optimal image cropping area according to the cropping judgment criteria. The adaptive cropping solution module uses the built-in video image stabilization unit to perform time-series calibration on the cropping position of short videos and live streaming frames to unify the inter-frame cropping position. S5. Through the composition optimization and output module with a multi-level serial unit structure, the cropped image is sequentially repaired for defects, corrected for proportions, and fine-tuned for composition, outputting short video and live media images that are adapted to various terminal proportions, have complete core content, and have a balanced composition.

[0008] Preferably, the media material acquisition and pattern recognition module in step S1 consists of an environment recognition unit and a dual-mode scheduling unit; the environment recognition unit is used to distinguish the operating environment of cloud devices and terminal devices, and the dual-mode scheduling unit is respectively connected to the cloud offline batch processing model and the device real-time lightweight model to form a dual-model association switching structure.

[0009] Preferably, the dual-branch deep feature extraction module in step S2 is an independent dual-path parallel fixed architecture, consisting of a key content recognition unit, an aesthetic composition evaluation unit, and a post-matching feature alignment unit that do not interfere with each other; the two recognition units synchronously and in parallel output two types of independent features, and the feature alignment unit unifies the dual-path feature specifications.

[0010] Preferably, the dynamic weight fusion module in step S3 consists of a screen type recognition unit and a weight adaptive adjustment unit; the screen type recognition unit is pre-connected to the dual-branch deep feature extraction module to distinguish between the main screen and the regular scene screen; the weight adaptive adjustment unit adaptively matches the dual feature ratio according to the screen type.

[0011] Preferably, the adaptive cropping solution module in step S4 consists of a proportional adaptation unit, a window traversal calculation unit, and an optimal filtering unit that are linked in sequence; the proportional adaptation unit supports arbitrary display ratio adaptation without fixed specifications, and the window traversal calculation unit and the optimal filtering unit are linked to match the optimal cropping area.

[0012] Preferably, the adaptive cropping solution module in step S4 has a built-in integrated video stabilization unit; the video stabilization unit consists of a position detection subunit and a smoothing correction subunit, forming an inter-frame temporal calibration closed-loop structure.

[0013] Preferably, the composition optimization and output module in step S5 is a three-level fixed series architecture, which is configured with an edge repair unit, a scale correction unit, and a center of gravity fine-tuning unit in sequence, with each unit linked and executed sequentially.

[0014] Preferably, the method used to execute S1-S5 is composed of a media material acquisition and pattern recognition module, a dual-branch deep feature extraction module, a dynamic weight fusion module, an adaptive cropping solution module, and a composition optimization and output module, which are sequentially coupled in series to form a fixed pre-input-feature extraction-weight fusion-cropping solution-post-optimization closed-loop processing architecture.

[0015] Preferably, the system is a heterogeneous layered deployment architecture between the cloud and the terminal, which is divided into a cloud deployment layer and a terminal deployment layer. The cloud deployment layer is configured with a complete model and a batch scheduling unit, while the terminal deployment layer is configured with a lightweight model and a real-time inference unit. The two-layer architecture has a division of labor and collaboration and supports synchronous iteration of model parameters.

[0016] This invention provides a deep learning-based adaptive intelligent cropping and composition optimization method for digital media content. It offers the following advantages: (1) This solution uses the dual-branch deep feature extraction module and the dynamic weight fusion module to collect key content information and aesthetic layout information of the image simultaneously. It can flexibly adjust the proportion of the two types of information in the cropping judgment according to the actual content of the image, avoid the problem that the traditional center cropping method is easy to crop out the core elements of the image, take into account the integrity of the material content and the composition effect of the image, and break through the technical limitation of the traditional cropping structure that is difficult to achieve a balance between content retention and visual aesthetics.

[0017] (2) Based on the internal unit cooperation relationship of the media material acquisition and pattern recognition module, the whole system can be adapted to two operating modes: cloud batch offline processing and terminal real-time online processing. Based on the end-cloud layered deployment structure, the model is configured hierarchically. It can adapt to different application scenarios such as short video batch production and live real-time image editing without the need for two separate processing devices. This optimizes the system deployment structure and expands the applicability of the equipment.

[0018] (3) A video image stabilization unit is added inside the adaptive cropping solution module. At the same time, a multi-level series arrangement of composition optimization and output modules can be used to perform linkage calibration on the cropping position of continuous video frames. After cropping, the image defect repair, size calibration and composition fine adjustment are completed in sequence. This improves the problem of image shaking and insufficient regularity of finished image after cropping of dynamic materials, reduces the workload of subsequent manual repair, and improves the whole process optimization system of dynamic media materials. Attached Figure Description

[0019] Figure 1 This is a block diagram of the overall system module serial architecture of the present invention; Figure 2 This is a schematic diagram of the internal unit structure of the dual-branch depth feature extraction module of the present invention; Figure 3 This is a schematic diagram of the internal unit structure of the adaptive trimming solution module of the present invention; Figure 4 This is a schematic diagram of the internal unit structure of the composition optimization and output module of the present invention; Figure 5 This is a flowchart illustrating the overall implementation process of the method of this invention.

[0020] In the diagram, 1. Media material acquisition and pattern recognition module; 11. Environment recognition unit; 12. Dual-mode scheduling unit; 2. Dual-branch deep feature extraction module; 21. Key content recognition unit; 22. Aesthetic composition evaluation unit; 23. Feature alignment unit; 3. Dynamic weight fusion module; 31. Screen type recognition unit; 32. Weight adaptive adjustment unit; 4. Adaptive cropping solution module; 41. Proportion adaptation unit; 42. Window traversal operation unit; 43. Optimal selection unit; 44. Video image stabilization unit; 441. Position detection subunit; 442. Smoothing correction subunit; 5. Composition optimization and output module; 51. Edge repair unit; 52. Proportion correction unit; 53. Center of gravity fine-tuning unit. Detailed Implementation

[0021] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention. Example 1

[0022] Cloud-based deployment and centralized optimization of large batches of short video materials by MCN agencies (see Figures 1, 2, 3, 4, and 5). This embodiment is deployed on a cloud-based multi-GPU computing server cluster, and is designed for the centralized and automated cropping and processing of massive amounts of finished video materials and original shooting frames for short video production companies. The system is configured with a deep learning model with complete weights and relies on the edge-cloud layered architecture of this invention to achieve batch offline computing.

[0023] After the device is powered on, the media material acquisition and pattern recognition module 1 connects to the enterprise's distributed material storage. The internal environment recognition unit 11 collects server computing power, hardware architecture and other operating parameters in real time. Through built-in discrimination logic, it determines that the current environment is a high-computing cloud environment. Then, the dual-mode scheduling unit 12 automatically connects to the full version of the neural network model and locks the offline batch processing mode. The raw material data is sent down to the dual-branch deep feature extraction module 2. This module adopts a physically parallel dual-path architecture, and the two paths are independent and do not interfere with each other. The key content recognition unit 21 relies on the pre-trained neural network to lock the key targets that need to be forcibly retained in the picture, such as the portrait subject, subtitle characters, and product logo, and generates position feature data. The aesthetic composition evaluation unit 22 simultaneously extracts layout features such as the size of the white space, the center of gravity of the scene, and the symmetry of the picture from the other path. The two feature parameters have different dimensions and formats. The feature alignment unit 23 set in the back end completes the data dimension unification and feature matching calibration to ensure that the two types of features can participate in the subsequent weight calculation simultaneously.

[0024] After the data enters the dynamic weight fusion module 3, the image type recognition unit 31 distinguishes between close-up frames and empty distant scene frames based on feature information. The weight adaptive adjustment unit 32 dynamically adjusts the weight ratio of the two types of features in the cropping judgment according to the image classification results, generating an independent cropping judgment benchmark adapted to a single frame. The internal ratio adaptation unit 41 of the adaptive cropping solution module 4 pre-locks the common 9:16 output ratio of the short video platform. The window traversal calculation unit 42 traverses the coordinates of all feasible cropping windows within the entire image range. The optimal selection unit 43 combines the aforementioned judgment benchmark to select the optimal cropping area from all windows. For continuous short video sequence frames, the module's built-in video image stabilization unit 44 starts the timing constraint logic. The position detection subunit 441 records the optimal cropping coordinates of adjacent frames in real time. The smoothing correction subunit 442 performs mean smoothing on the coordinate abrupt changes to limit large jumps in the cropping box between frames. The cropped image data flows into the composition optimization and output module 5. Under the sequential processing of the three-level series units of edge repair unit 51, proportion correction unit 52, and center of gravity fine adjustment unit 53, the edge defects caused by cropping are repaired, the target size is accurately matched, and the center of gravity of the image is finely adjusted. The finished material is automatically classified and stored in the cloud finished product database after processing.

[0025] After comparing and testing tens of thousands of materials, compared with the traditional center-fixed cutting process, the error rate of key content in this embodiment is significantly reduced, and the overall automation processing efficiency is significantly improved. Example 2

[0026] Mobile hardware embedding, real-time online cropping and optimization of e-commerce live streaming footage (refer to Figures 1, 2, 3, 4, and 5). This embodiment embeds the entire system into the motherboard of a smartphone live streaming platform, adopts a lightweight and compressed inference architecture, and is adapted to the entire application chain of real-time image acquisition, cropping, and streaming for e-commerce live streaming. It relies on terminal hardware to achieve low-latency online computing.

[0027] After the device starts streaming, the media material acquisition and pattern recognition module 1 captures the original video bitstream output by the camera in real time. The environment recognition unit 11 collects the terminal CPU and memory hardware parameters and identifies that the current mobile terminal is a low computing power hardware environment. The dual-mode scheduling unit 12 automatically switches to the lightweight inference model and starts the low-latency real-time processing mode.

[0028] Real-time video data is continuously fed into the dual-branch deep feature extraction module 2. The two parallel paths run synchronously. The key content recognition unit 21 extracts key target features such as the anchor's portrait, shopping pop-up text, and product subject in real time. The aesthetic composition evaluation unit 22 collects composition parameters such as the layout of the live room scene and the white space. The two features are sent to the downstream module after being unified in data format by the feature alignment unit 23.

[0029] The dynamic weight fusion module 3 relies on the image type recognition unit 31 to judge changes in image content in real time. When the image focuses on the anchor and the products for sale, the weight adaptive adjustment unit 32 increases the weight of key content features. When the camera switches to an outdoor scene or a distant view of the product, the weight of aesthetic composition features is automatically increased, and real-time cropping criteria are dynamically generated. The adaptive cropping solution module 4 allows the broadcaster to customize the output image ratio before broadcasting. The ratio adaptation unit 41 dynamically constrains the cropping boundary range according to the set parameters, and the window traversal calculation unit 42 and the optimal filtering unit 43 calculate the cropping area in real time. The video image stabilization unit 44 keeps the inter-frame constraint logic enabled throughout the broadcast, continuously comparing the cropping coordinates of the previous and next frames. Once the coordinate fluctuation exceeds the preset threshold, the smoothing correction subunit 442 makes a small correction to the coordinates to eliminate the frequent shaking and flickering of the live broadcast image.

[0030] After cropping and optimization, the image undergoes step-by-step edge flaw repair, image size calibration, and center of gravity fine-tuning in the composition optimization and output module 5. The final processed video data is then directly connected to the streaming encoding interface to transmit live signals.

[0031] This embodiment operates fully automatically, requiring no operator to drag and drop the cropping frame in real time, thus meeting the stringent requirement of millisecond-level low-latency output in live streaming scenarios. Example 3

[0032] Edge-cloud collaborative architecture; integrated processing of multiple specifications of mixed materials including static images and short videos (refer to Figures 1 and 2). Figure 3 (Figures 4 and 5) This embodiment establishes a cloud + local terminal collaborative operation architecture, which is applied to the material processing system of self-media content platform. It can process two types of materials at the same time: product promotional static images and daily short videos, and supports three commonly used output ratios of new media platforms: 1:1, 3:4, and 9:16.

[0033] After the system went live, the media material acquisition and pattern recognition module 1 uniformly received all types of materials uploaded by users in batches. Relying on the environment recognition unit 11 to identify idle computing resources, the dual-mode scheduling unit 12 achieved intelligent distribution: large batches of static poster images were allocated to the cloud server for offline centralized computation using a complete model; scattered short video clips were allocated to local terminals for real-time online processing using a lightweight model. All materials after distribution were uniformly sent to the dual-branch deep feature extraction module 2. The module's dual-path parallel architecture does not distinguish between static images and dynamic videos, and uniformly completes the parallel extraction of key content features and aesthetic composition features. The feature alignment unit 23 unifies the data specifications of the two types of features, eliminating data format differences caused by different materials.

[0034] The dynamic weight fusion module 3 adaptively adjusts the weight ratio based on the main subject distribution of each frame, regardless of the static or dynamic attributes of the material, and independently generates the corresponding cropping standard. The adaptive cropping solution module 4's proportion adaptation unit 41 flexibly switches between three preset output proportions based on the target publishing platform of the material. The system has built-in intelligent control logic that only activates the video stabilization unit 44 to perform adjacent frame coordinate calibration processing for short video materials; static single images do not require timing correction and skip this unit process directly. All cropped images and videos are uniformly entered into the composition optimization and output module 5, where edge repair, size correction, and center of gravity optimization are completed sequentially through three levels of interconnected units. The system automatically creates categorized folders according to material type and output aspect ratio and archives the finished products.

[0035] This embodiment relies on only one set of hardware and software architecture to be compatible with the processing needs of multiple types of materials and multiple output specifications, avoiding the need for staff to switch back and forth between multiple cutting software, and effectively reducing the labor and time costs of post-processing of materials.

[0036] Working principle: The overall system of this invention consists of a media material acquisition and pattern recognition module 1, a dual-branch deep feature extraction module 2, a dynamic weight fusion module 3, an adaptive cropping solution module 4, and a composition optimization and output module 5, which are sequentially coupled to form a fully closed-loop processing architecture. The system adopts a heterogeneous layered deployment architecture between the edge and cloud. Relying on the collaborative cooperation of the internal units of each module, it completes the entire process of short video frames, live real-time images, and static images from material acquisition to cropping and optimization output. The overall workflow is implemented step by step in five stages: pattern determination and scheduling, parallel extraction of dual-dimensional features, adaptive weight fusion, intelligent cropping solution, and post-integrated composition optimization. The specific principle is as follows: Intelligent determination of media access and operating mode. The media acquisition and pattern recognition module 1 is responsible for the unified collection of various original media media. The internal environment recognition unit 11 collects the device's computing power, hardware configuration, and operating environment parameters in real time, and automatically distinguishes between the high computing power environment of the cloud server and the low computing power environment of the mobile terminal. The dual-mode scheduling unit 12 connects to the complete cloud processing model or the lightweight terminal model according to the environment recognition results, and switches between the cloud batch offline processing mode and the terminal real-time online processing mode accordingly to complete the media distribution and system operating condition configuration, so as to achieve a hardware architecture that is compatible with two working scenarios from the source.

[0037] Dual-path parallel feature extraction and data regularization. The source data is fed into the dual-branch deep feature extraction module 2, which uses two independent parallel computing paths. The key content recognition unit 21 uses a deep learning network to locate key targets in the image that must be retained, such as portraits, subtitles, and trademarks, and outputs key position feature data. The aesthetic composition evaluation unit 22 extracts layout feature data such as the proportion of white space, the center of gravity of objects, and the symmetry of the image from another independent path. Since the data dimensions and data formats of the two types of features are inconsistent, the feature alignment unit 23 set up afterward unifies the specifications and dimensions of the two types of features, so that the key content features and aesthetic composition features can achieve data matching, providing a compliant data source for subsequent weight fusion, which is different from the working logic of single feature extraction in existing technologies.

[0038] Image classification and dynamic weight adaptive fusion. The normalized feature data is input into the dynamic weight fusion module 3. The image type recognition unit 31 automatically distinguishes between close-up shots of the main subject and empty shots of distant scenery based on the dual-dimensional features transmitted from the front end. The weight adaptive adjustment unit 32 dynamically adjusts the weight ratio of key content features and aesthetic composition features in the cropping judgment according to the image classification results. The weight of key content features is increased for the main subject and the weight of aesthetic composition features is increased for empty shots. Based on the dynamic ratio relationship, a special cropping judgment standard adapted to a single frame image is generated, which abandons the inherent defects of the fixed weight uniform calculation of the traditional solution.

[0039] Multi-ratio adaptive cropping and dynamic video frame stabilization control. The judgment criteria are input into the adaptive cropping solution module 4. The ratio adaptation unit 41 supports any customizable output ratio, unrestricted by fixed size. The window traversal calculation unit 42 traverses all feasible cropping window coordinates. The optimal selection unit 43, combined with the cropping criteria generated in the previous step, selects the optimal cropping boundary from all windows. For short videos, live streams, and other continuous frame dynamic materials, the module's built-in video stabilization unit 44 initiates timing constraint logic. The position detection subunit 441 records the optimal cropping coordinates of adjacent frames in real time and calculates the coordinate deviation. The smoothing correction subunit 442 performs smoothing averaging on coordinates with deviations exceeding a threshold, constraining large jumps in the cropping frames of adjacent frames. Static images without continuous frame data automatically skip the inter-frame calibration process.

[0040] The three-stage sequential post-processing image optimization and output: The cropped images are uniformly sent to the composition optimization and output module 5. The module adopts a three-stage sequential architecture consisting of an edge repair unit 51, a proportion correction unit 52, and a center of gravity fine-tuning unit 53. The images are processed step by step in a front-to-back order: the edge repair unit repairs the edge defects caused by cropping, the proportion correction unit accurately matches the preset output size, and the center of gravity fine-tuning unit corrects the image's center of gravity shift. After the three-stage optimization process, the output is a finished media material with a neat composition and complete key content.

[0041] In addition, the system of this invention relies on a heterogeneous deployment architecture between the cloud and the edge. The cloud deployment layer is equipped with a complete model and a batch scheduling unit, which is responsible for the offline centralized calculation of large amounts of materials. The terminal deployment layer is equipped with a lightweight model and a real-time inference unit, which is responsible for the real-time processing of small amounts of materials. The cloud and the terminal support synchronous iterative updates of model parameters, which further ensures the recognition and cropping accuracy of the system during long-term use.

[0042] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0043] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A deep learning-based adaptive intelligent cropping and composition optimization method for digital media content, characterized in that, The steps are as follows: S1. The media material acquisition and pattern recognition module (1) collects original media materials such as short video frames, live real-time images, and static images, identifies the current device operating environment, and adaptively switches between cloud batch processing mode and device real-time processing mode according to the differences in device environment. S2. Through the dual-branch deep feature extraction module (2), a dual-path parallel structure is adopted to simultaneously and independently extract key content features and aesthetic composition features. The key content features are used to locate the core target area that needs to be retained in the picture. The aesthetic composition features are used to characterize the visual balance and layout of the picture. S3. By associating the output features of the dual-branch deep feature extraction module with the dynamic weight fusion module (3), the two types of features are matched according to the differences in the content type of the screen, and a special cropping judgment standard adapted to the current screen is generated. S4. By calling the preset ratio adaptation structure and the optimal solution structure through the adaptive cropping solution module (4), the screen display ratio can be adapted to any custom screen display ratio, and the optimal screen cropping area can be locked according to the cropping judgment standard. Through the video screen stabilization unit (44) built into the adaptive cropping solution module (4), the cropping position of short video and live continuous frames is time-ordered and the inter-frame cropping position is unified. S5. Through the composition optimization and output module (5) with a multi-level serial unit structure, the cropped image is sequentially repaired for defects, corrected for proportions, and fine-tuned for composition, and outputs short video and live media images that are adapted to various terminal proportions, have complete core content, and have balanced composition.

2. The method for intelligent cropping and composition optimization of adaptive digital media content based on deep learning according to claim 1, characterized in that: The media material acquisition and pattern recognition module (1) mentioned in step S1 is composed of an environment recognition unit (11) and a dual-mode scheduling unit (12). The environment recognition unit (11) is used to distinguish the operating environment of cloud devices and terminal devices. The dual-mode scheduling unit (12) is connected to the cloud offline batch processing model and the device real-time lightweight model respectively, forming a dual-model association switching structure.

3. The method for adaptive intelligent cropping and composition optimization of digital media content based on deep learning according to claim 2, characterized in that: The dual-branch deep feature extraction module (2) in step S2 is an independent dual-path parallel fixed architecture, consisting of a key content recognition unit (21) that does not interfere with each other, an aesthetic composition evaluation unit (22), and a post-matching feature alignment unit (23); the two recognition units synchronously and in parallel output two types of independent features, and the feature alignment unit (23) unifies the dual-path feature specifications.

4. The method for adaptive intelligent cropping and composition optimization of digital media content based on deep learning according to claim 3, characterized in that: The dynamic weight fusion module (3) in step S3 consists of a screen type recognition unit (31) and a weight adaptive adjustment unit (32); the screen type recognition unit (31) is connected to the dual-branch deep feature extraction module (2) in advance to distinguish between the main screen and the regular scene screen; the weight adaptive adjustment unit (32) adaptively matches the dual feature ratio according to the screen type.

5. The method for adaptive intelligent cropping and composition optimization of digital media content based on deep learning according to claim 4, characterized in that: The adaptive cropping solution module (4) in step S4 consists of a proportional adaptation unit (41), a window traversal calculation unit (42), and an optimal filtering unit (43) linked in sequence. The proportional adaptation unit (41) supports arbitrary display ratio adaptation without fixed specifications, and the window traversal calculation unit (42) and the optimal filtering unit (43) are linked to match the optimal cropping area.

6. The method for adaptive intelligent cropping and composition optimization of digital media content based on deep learning according to claim 5, characterized in that: The adaptive cropping solution module (4) in step S4 has a built-in integrated video image stabilization unit (44); the video image stabilization unit (44) consists of a position detection subunit (441) and a smoothing correction subunit (442), forming an inter-frame temporal calibration closed-loop structure.

7. The method for adaptive intelligent cropping and composition optimization of digital media content based on deep learning according to claim 6, characterized in that: The composition optimization and output module (5) mentioned in step S5 is a three-level fixed series architecture, which is configured with edge repair unit (51), proportion correction unit (52) and center of gravity fine adjustment unit (53) in sequence. Each unit is linked and executed sequentially.

8. The method for adaptive intelligent cropping and composition optimization of digital media content based on deep learning according to claim 7, characterized in that: The method used to execute S1-S5 is composed of a media material acquisition and pattern recognition module (1), a dual-branch deep feature extraction module (2), a dynamic weight fusion module (3), an adaptive cropping solution module (4), and a composition optimization and output module (5) connected in series and coupled to form a fixed pre-input-feature extraction-weight fusion-cropping solution-post-optimization closed-loop processing architecture.

9. The method for adaptive intelligent cropping and composition optimization of digital media content based on deep learning according to claim 8, characterized in that: The system adopts a heterogeneous layered deployment architecture, consisting of a cloud deployment layer and a terminal deployment layer. The cloud deployment layer is configured with a complete model and a batch scheduling unit, while the terminal deployment layer is configured with a lightweight model and a real-time inference unit. The two layers work together and support synchronous iteration of model parameters.

Images