|
List of Papers:
Editorial of this Book
AOI :10.100.232025.00101
Click to show/hide Abstract:
ABSTRACT:
Low contrast and poor illumination limit how well land cover, roads and settlements can be interpreted in optical remote sensing images. Wavelet–SVD methods tuned by nature-inspired algorithms (NIA) such as PSO and GWO are fast and interpretable, but they search per image for a proxy fitness (mean-intensity matching). We show that their SVD correction reduces to a single scalar gain on the approximation band, and that the detail bands are left untouched, so weak detail is attenuated relative to the brightened background. We propose AI-DWTE, which replaces the per-image search with a learned predictor. A deep ensemble of multilayer perceptrons maps 100 wavelet-domain statistics to the eight parameters of a Haar-DWT operator in a single forward pass. The operator combines black-point removal, channel-wise power-law mapping, a learned singular-value gain, local contrast, and a new gain-consistent detail transfer that scales each detail coefficient by the local slope of the tone curve after wavelet shrinkage. The network is trained on parameters found by an offline oracle that has access to ground truth. On 420 held-out UC Merced images under three degradations (1,260 cases), AI-DWTE is compared against ten techniques, from HE and CLAHE through the PSO/GWO base method to the pretrained deep models Zero-DCE, Zero-DCE++ and SCI. It improves average PSNR from 17.95 dB (PSO base method) to 21.71 dB, SSIM from 0.787 to 0.851, and CIEDE2000 colour error from 13.85 to 9.75. It runs in 27 ms per image on one CPU thread, 4.3× faster than the PSO search. Under an unseen non-uniform illumination degradation, its gain over the base method is small (+0.15 dB) and not statistically significant, a limitation we report and analyse.
AI-DWTE: Learning-Guided Wavelet-Domain Enhancement of Low-Contrast Remote Sensing Images with Gain-Consistent Detail Preservation
Rajni Sharma, Ravinder M
Dept. of CSE, Indira Gandhi Delhi Technical University for Women, Kashmere Gate, New Delhi, India.
--------------------------------------------------------------------------------------
|
Download Paper template of ACT Proceedings from this link.
---------------------------------------------------------------------------------------------
Pay to book project: from India pay by scanning QR code:
Download Per paper, Pay $10 (USD)(or equivalent)(from outside of India)
From outside India:pay through our PayPal account
Account details:
Account Name: Applied Computer Technology
Account Number: 35532552072
Bank name: STATE BANK OF INDIA
IFSC Code: SBIN0001404
Account Type: Current
Branch Name: Kamarhati
Address: 1, B.T.Road, Kolkata-700058, West Bengal, India.
PAN : AVPPA7870E
Phone : +91-8420582707
--------------------------------------
For query about publishing for printed or online version of conference Proceedings, write to: info@actsoft.org or
talk to: +91-6291839750
About indexing of AOI(Applied Object Indexing): it is a new and advanced indexing method for digital and real life hard objects. Any book, paper, picture, medicine,
Land, Building etc. can be indexed with unique number. Clicking that number, the object can be identified with its contents.
List of all papers
---------------------------------------------------------------------------------------------
|
N.B. After publication of all individual papers, full book will be available as PDF format for downloading. *** Any Conference organizers can write to: info@actsoft.org for publishing papers in ACT Proceedings ***