Search

CN-121998915-A - Processing system for potential representation of image

CN121998915ACN 121998915 ACN121998915 ACN 121998915ACN-121998915-A

Abstract

The invention discloses a processing system for potential representation of an image, which comprises a potential representation verification unit 1, a super prior information generation unit 2 and a super prior quantization unit 3, wherein the potential representation verification unit 1 is used for evaluating the quality of the potential representation, the potential representation is restored into the image through inverse transformation and is compared with an original image, the super prior information generation unit 2 selects the most suitable method according to performance evaluation to generate the super prior information, and the super prior quantization unit 3 determines the most suitable quantization method according to the super prior information generated by the super prior information generation unit 2. The invention solves the defects in the prior art, and achieves remarkable technical effects in the aspects of potential representation verification, super priori information generation, quantization and the like. The technical effects not only improve the accuracy and reliability of the potential representation processing of the image, but also reduce the storage and transmission cost of the data, and provide new ideas and schemes for the development and application of the image processing technology.

Inventors

  • MENG LI
  • YANG SHUANG

Assignees

  • 西南交通大学

Dates

Publication Date
20260508
Application Date
20260107

Claims (10)

  1. 1. A processing system for potential representation of an image is characterized by comprising a potential representation verification unit (1), a super prior information generation unit (2) and a super prior quantization unit (3), wherein the potential representation verification unit (1) is used for evaluating quality of the potential representation, the potential representation is restored to the image through inverse transformation and is compared with an original image, whether key information of the original image is reserved in the potential representation is judged from the view point of the image, the super prior information generation unit (2) is used for selecting the most suitable method to generate the super prior information according to performance evaluation based on the potential representation verified by the potential representation verification unit (1) so as to extract key features in the potential representation and conduct dimension compression, and the super prior quantization unit (3) is used for determining the most suitable quantization method according to the super prior information generated by the super prior information generation unit (2) and setting quantization parameters through parameter analysis and experiment optimization so as to realize quantization of the super prior information.
  2. 2. A processing system for potential representations of images according to claim 1, characterized in that the potential representation verification unit (1) further comprises an image inverse transformation unit (11) and an image comparison unit (12), the image inverse transformation unit (11) being arranged to restore the potential representation to an image by means of a preset inverse transformation algorithm, the image comparison unit (12) being arranged to compare the image resulting from the inverse transformation with the original image.
  3. 3. The processing system for potential representation of image according to claim 2, wherein the super prior information generating unit (2) further comprises an analysis method researching unit (21), a performance evaluating unit (22) and a method selecting unit (23), wherein the analysis method researching unit (21) researches a plurality of analysis methods such as statistical analysis, machine learning and deep learning, the performance evaluating unit (22) performs performance evaluation from a plurality of dimensions such as computational complexity, feature extraction capability and adaptability to potential representation data for different analysis methods, and the method selecting unit (23) selects the most suitable analysis method to generate the super prior information according to the evaluation result of the performance evaluating unit (22).
  4. 4. A processing system for potential representation of an image according to claim 3, wherein the super prior quantization module (3) further comprises a quantization method investigation unit (31), a quantization parameter determination unit (31) and a quantization error processing unit (33), wherein the quantization method investigation unit (31) is used for investigating different quantization methods, including scalar quantization and vector quantization, the quantization parameter determination unit (32) is used for determining quantization parameters through parameter analysis and experimental optimization, and the quantization error processing unit (33) is used for processing quantization errors which are inevitably generated in the quantization process by adopting various strategies.
  5. 5. A processing system for a potential representation of an image as recited in claim 4, wherein said plurality of strategies includes a prediction-based error compensation strategy and a noise shaping strategy.
  6. 6. The processing system for potential representation of an image of claim 5, wherein the image comparison unit (12) in the potential representation verification module further comprises a multi-scale comparison subunit (121) and a structural similarity evaluation subunit (122), the multi-scale comparison subunit (121) compares the inversely transformed image with the original image from different scales, and the structural similarity evaluation subunit (122) evaluates the similarity of the inversely transformed image with the original image using a structural similarity index.
  7. 7. A processing system for a potential representation of an image according to claim 6, wherein said potential representation verification unit further comprises a feature matching unit (13), the feature matching unit (13) further comprising a feature point detection sub-unit (131), a feature descriptor generation sub-unit (132) and a matching relation determination sub-unit (133).
  8. 8. A processing system for potential representation of an image according to claim 7, wherein the feature point detection subunit (131) employs a plurality of feature point detection algorithms to detect feature points in the inversely transformed image and the original image, the feature descriptor generation subunit (132) generates feature descriptors for the detected feature points, and the matching relationship determination subunit (133) determines a matching relationship between the inversely transformed image and the original image feature points using the generated feature descriptors.
  9. 9. A processing system for potential representation of an image according to claim 8, wherein said super prior information generating unit (2) further comprises a super prior information fusion unit (24) for fusing super prior information of different sources or different types.
  10. 10. A processing system for potential representations of images according to claim 9, characterized in that the image inverse transformation unit (11) comprises an inverse transformation network (111), a generator part of the inverse transformation network (111) adopts a multi-layer convolution transposed neural network structure, potential representation vectors are mapped to a low-resolution feature map through a fully connected layer from the potential representation as input, the feature map enters the convolution transpose layers, each convolution transpose layer consists of convolution transpose operation, batch normalization operation and ReLU activation function, and a discriminator of the inverse transformation network (111) adopts the multi-layer convolution neural network structure.

Description

Processing system for potential representation of image Technical Field The invention belongs to the technical field of image processing, and particularly relates to a processing system aiming at potential representation of an image. Background The image latent representation (Latent Representation of Images) refers to the feature representation obtained by converting the image data into a low-dimensional latent space (LATENT SPACE) through a specific transformation or mapping. In the field of data processing and analysis today, potential representations are widely used as a way to abstract and compress the original image data. The potential representation can reduce the dimension of the data while retaining the key characteristics of the data, thereby improving the efficiency of data processing, and plays an important role in various fields such as image recognition, voice processing and the like. However, there are some disadvantages to quality assessment of potential representations and generating and quantifying super prior information based on the potential representations. In terms of potential representation verification, the existing verification method is mostly limited to simple image restoration comparison, and the single verification method cannot comprehensively and deeply evaluate the quality of the potential representation, so that some important semantic information or multi-mode information loss conditions in the potential representation can be omitted. In the generation link of the super prior information, the existing method often adopts a single analysis means, and lacks of system comparison and on-demand selection of multiple analysis methods. For example, using only simple statistical analysis methods may not fully mine complex features in potential representations, while relying solely on deep learning methods may face problems of high computational complexity, difficult model training, and the like. Meanwhile, a mechanism for dynamically adjusting an analysis method is lacking, and characteristics potentially representing real-time changes of data cannot be adapted. In the aspect of super priori quantization, the current quantization method for super priori information is blind, and the characteristics of multiple relevant dimensions and the like of the super priori information are not fully considered, so that the quantization effect is poor. In addition, the setting of the quantization parameter is often based on an empirical value, and a parameter analysis and experimental optimization process of a system is lacked, so that quantization errors are larger, the data compression ratio is not ideal, and the requirements of data accuracy and high-efficiency storage and transmission in practical application cannot be met. In view of the foregoing, there is a need for a more comprehensive, systematic and efficient solution to these problems in all key links of the existing potential representation processing technology. Disclosure of Invention The present invention is to solve the following technical problems. ‌, Potentially representing technical problems in verification. The existing potential representation verification mode is mostly limited to simple image restoration contrast, and has obvious limitation. First, it cannot fully and deeply evaluate the quality of the potential representation. The potential representation may contain abundant semantic information and multi-modal information, and a simple image restoration comparison often only reveals the retention of part of visual features, and the loss of other key information may not be effectively detected. Second, a single verification method may be affected by subjective factors, resulting in questioning the accuracy and reliability of the evaluation result. Therefore, the technical problem to be solved in the aspect of potential representation verification is to provide a more comprehensive and objective verification method, which can accurately evaluate the quality of potential representation and reveal key information possibly lost in the potential representation. ‌ Two, technical problems in the aspect of super priori information generation. In the generation link of the super prior information, the existing method often adopts a single analysis means, and lacks of system comparison and on-demand selection of multiple analysis methods. A limitation of this approach is that different analysis methods may be suitable for different types of potential representation data, while a single analysis method may not adequately mine complex features in the potential representation. Furthermore, the lack of a mechanism to dynamically adjust the analysis method also makes the existing methods unable to accommodate the real-time changing nature of the underlying presentation data. Therefore, the technical problem to be solved in the aspect of super prior information generation is to provide a method capable of systematically comparing an