AKKUL ERDEM SAFA (TR)
ÇARKACIOĞLU LEVENT (TR)
TÖREYİN BEHÇET UĞUR (TR)
UNIV ISTANBUL TEKNIK (TR)
| CN107292256A | 2017-10-24 | |||
| CN102867187A | 2013-01-09 | |||
| CN103207999A | 2013-07-17 |
TANG JIEXIONG ET AL: "Compressed-Domain Ship Detection on Spaceborne Optical Image Using Deep Neural Network and Extreme Learning Machine", IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, IEEE, USA, vol. 53, no. 3, 1 March 2015 (2015-03-01), pages 1174 - 1185, XP011559177, ISSN: 0196-2892, [retrieved on 20140916], DOI: 10.1109/TGRS.2014.2335751
LI X ET AL: "Detecting faces in the wavelet compressed domain", VISUAL COMMUNICATIONS AND IMAGE PROCESSING; 12-7-2005 - 15-7-2005; BEIJING,, 12 July 2005 (2005-07-12), XP030080922
TAO GUIHUA ET AL: "Incorporating Discrete Wavelet Transformation Decomposition Convolution into Deep Network to Achieve Light Training", 7 September 2021, TOPICS IN CRYPTOLOGY - CT-RSA 2020 : THE CRYPTOGRAPHERS' TRACK AT THE RSA CONFERENCE 2020, SAN FRANCISCO, CA, USA, FEBRUARY 24-28, 2020, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, PAGE(S) 307 - 318, XP047607137
SALIM LAHMIRI ET AL: "Classification of brain MRI using the LH and HL wavelet transform sub-bands", CIRCUITS AND SYSTEMS (ISCAS), 2011 IEEE INTERNATIONAL SYMPOSIUM ON, IEEE, 15 May 2011 (2011-05-15), pages 1025 - 1028, XP031997806, ISBN: 978-1-4244-9473-6, DOI: 10.1109/ISCAS.2011.5937743
YAO-HONG TSAI ET AL: "Wavelet-Based Image Fusion by Adaptive Decomposition", INTELLIGENT SYSTEMS DESIGN AND APPLICATIONS, 2008. ISDA '08. EIGHTH INTERNATIONAL CONFERENCE ON, IEEE, PISCATAWAY, NJ, USA, 26 November 2008 (2008-11-26), pages 283 - 287, XP031368505, ISBN: 978-0-7695-3382-7
| CLAIMS 1. A method that reduces the performance degradation in JPEG2000 compressed images and provides more efficiency with the benefits of partial decompression method (10), and is characterized in that it comprises the process steps of: • Partial decompressing of images compressed with the JPEG2000 algorithm to access their sub-bands at 4 times the original size and obtaining XLL, XLH, XHL and XHH, • With Hybrid Maximum Sub-Band Method (31): > Taking the absolute value of the pixel values in the XLH, XHL and XHH bands and projecting the distribution of pixel values in these bands as integers between [0,+128], > Forming the XMAX by combining the XLH, XHL and XHH bands with the maximum pixel retrieval process, > Dividing all elements of the XMAX matrix by 128 and having all pixel values as 32-bit double values distributed as real numbers between [0,1], > Dividing all elements of the XLL band by 256 and having all pixel values as 32-bit double values distributed as real numbers between [0,1], > Forming the HMAX by combining the XLL band and the XMAX band, > Converting the HMAX band to be represented by 8 bits with the following process, • With Hybrid Average Sub-Band Method (21): > Adding 128 to all pixel values of XLH , XHL and XHH bands and obtaining the distribution of pixel values in these bands as integers in the range [0,255], > Averaging the XLH , XHL and XHH bands and forming the XAVG, > Dividing all elements of the XAVG matrix by 255 and having all pixel values as 32-bit double values distributed as real numbers between [0,1], > Dividing all elements of the XLL band by 256 and having all pixel values as 32-bit double values distributed as real numbers between [0,1], > Forming the HAVG by combining the XLL band and the XAVG band, > Converting the HAVG band to be represented by 8 bits with the following process, |
Technical Field
The invention relates to a method for minimizing the performance degradation of JPEG2000 compressed images and maximizing the benefits of Partial Decompression.
State of the Art
Disadvantages of JPEG2000 compressed images are that they take up a lot of memory, decompression times are quite long after compression, and the sub-images used in partial decompression of JPEG2000 compressed images do not achieve sufficient performance during image analysis. Although the partial decompression approach for JPEG2000 compressed images provides speed and space savings, it is not preferred for critical architectures due to the decrease in performance. This results in the system running slower and requiring more memory, as well as the need for components with more processing capacity.
In the application numbered JP5175820B2, the sub-images of the images are extracted and the pre-evaluation pixel value of the original image is found. (This value represents the median of the pixel values that the image can take. For example, an 8-bit image is represented by 256 pixels, so the pre-evaluation pixel value will be 128.) When compressing images, all subimages are divided into 4x4 segments and the peak pixel in this segment is found. When calculating the peak pixel, the pixel average of the 4x4 partition is calculated. If this average is greater than the pre-evaluation value, the peak pixel is assigned as the largest pixel value of this segment. If it is below, the smallest pixel value is assigned as the peak pixel. The locations of the peak pixel values are divided into 3x3 areas. When images are compressed, they are stored in memory according to their peak pixel positions. Based on the peak pixel position, images are decompressed and faster decompression is achieved.
In the above-mentioned application, averaging is used to find the pixel with the largest absolute value in 4x4 sections. In the inventive method, unlike this application, the values of different sub-images at the same pixel positions are averaged and a high-frequency average sub-image is created.
Consequently, an improvement in the related technical field has become necessary due to the drawbacks mentioned above and the inadequacy of the existing solutions. Objective of the Invention
The invention aims to provide a structure with different technical features that brings a new decompression method in this field, unlike the structures used in the current technique.
The primary objective of the invention is to provide a method that minimizes the image analysis performance degradation that occurs during the partial decompression of JPEG2000 compressed images and maximizes the benefits of the partial decompression approach.
The objective of the invention is to provide a method that maximizes performance gains by blending the high-frequency HL, LH, and HH sub-bands in JPEG2000 compressed images with the LL sub-band and by including "maximization" and "averaging" processes.
In tests on landscape images obtained by remote sensing, the inventive method was applied to the landscape classification task with the DenseNet-121 architecture. A performance gain of approximately 2% was achieved compared to the architecture trained with only the LL subband.
Rather than increasing the resolution, the objective of the invention is to smooth the edgecorner information in the image by transferring the information contained in the high-frequency sub-images to the LL band.
The structural and characteristic features of the invention and all its advantages will be more clearly understood from the following figures and the detailed description with references to these figures. Therefore, the evaluation should be made with reference to these figures and the detailed description.
Description of the Figures
Figure 1, Illustration of the summary of the Hybrid Maximum Sub-band and Hybrid Average Sub-band methods.
The drawings are not necessarily to scale and details that are not necessary to understand the present invention may be omitted. Furthermore, elements that are at least substantially identical or have at least substantially identical functions are indicated by the same number.
Description of Part References
10. Partial decompression method
20. Average sub-band method
21 . Hybrid average sub-band method
30. Maximum sub-band method
31. Hybrid maximum sub-band method Abbreviations
X: Fully decompressed version of the image.
R: Number of lines of sub-band images.
C: Number of columns of sub-band images.
XLL: The partially decompressed Low-Low (LL) sub-band of the image. It is a 14 reduction of the original image. The pixel values of the image vary as integers between [0,255],
XLH: The partially decompressed Low-High (LH) sub-band of the image. This is an image that has been reduced by 14 of the original image and contains detail information only in the vertical axis. The pixel values of the image vary as integers between [-128, +128],
XHL: The partially decompressed High-Low (HL) sub-band of the image. This is an image that has been reduced by 14 of the original image and contains detail information only in the horizontal axis. The pixel values of the image vary as integers between [-128, +128],
XHH: The partially decompressed High-High (HH) sub-band of the image. It is an image that has been reduced by 14 of the original image and contains detail information only on the diagonal axis. The pixel values of the image vary as integers between [-128, +128],
XMAX: The high-band matrix created by the maximization process. Each element is represented by 32 bits.
XAVG: The high-band matrix created by the averaging process. Each element is represented by 32 bits.
HMAX : Hybrid maximum sub-band.
Min(HMAx) : The minimum value of the HMAX band.
Max(HMAx) : The maximum value of the HMAX band.
HAVG : Hybrid average sub-band.
Min(HAVG) : The minimum value of the HAVG band.
Max(HAvc) : The maximum value of the HAVG band.
HH : High-High
LH : Low-High
HL : High-Low
LL : Low-Low
Detailed Description of the Invention
In this detailed description, the preferred embodiments of the invention are described only for the purpose of better understanding the subject matter and without limitation.
The invention relates to a method for reducing the degradation in performance of JPEG2000 compressed images, while maximizing the benefits of the partial decompression method (10). In the inventive method, a new sub-band is created by blending HL, LH and HH sub-bands with the LL sub-band. Maximization and averaging processes are applied on HL, LH and HH sub-bands. The intermediate sub-bands resulting from the maximization and averaging processes are blended with the LL sub-band. The LL sub-band is updated to be represented by 8 bits and a hybrid sub-band with a size four times smaller than the original image is obtained. The resulting hybrid sub-image has higher performance than the LL sub-band itself when used in image analysis (object classification, object detection, object segmentation, etc.). The invention has been tested for classification on remote sensing landscape images and found to have about 2% higher performance than the LL sub-band.
In the inventive method, images compressed with the JPEG2000 algorithm are partially decompressed to access sub-bands four times smaller than the original size. This results in XLL, XLH, XHL and XHH. (LL band: Comprises low-frequency details. It is a low resolution image of the original image. LH band: Comprises high-frequency vertical details. HL band: Comprises high-frequency horizontal details. HH band: Comprises high-frequency diagonal details).
The following steps are then followed for the two different algorithms.
Hybrid Maximum Sub-Band Method (31)
1 . The absolute value of the pixel values in the XLH , XHL and XHH bands are taken. This way, the distribution of pixel values in these bands is reflected between [0,+128],
2. The XLH, XHL and XHH bands are combined with the maximum pixel extraction to create
3. All elements of the XMAX matrix are divided by 128. Thus, all pixel values are real numbers with 32-bit double values distributed between [0,1],
4. All elements of the XLL band are divided by 256. Thus, all pixel values are real numbers with 32-bit double values distributed between [0,1],
5. The XLL band and the XMAX band are summed as follows to form HMAX.
6. The HMAX band is converted to be represented by 8 bits with the following process. Hybrid Average Sub-Band Method (21)
1. 128 is added to all pixel values of the XLH, XHL and XHH bands. This results in an integer distribution of pixel values in these bands in the range [0,255],
2. The and bands are averaged and XAVG is generated as follows.
3. All elements of the XAVG matrix are divided by 255. This way all pixel values have 32 bit double values distributed as real numbers between [0,1],
4. All elements of the XLL band are divided by 256. This way all pixel values have 32 bit double values with a real number distribution between [0,1], 5. The XLL band and the XAVG band are summed as follows to form HAVG.
6. The HAVG band is converted to be represented by 8 bits with the following process.