LEE- Bilgisayar Mühendisliği-Yüksek Lisans

Bu koleksiyon için kalıcı URI

http://hdl.handle.net/11527/19207

Gözat

Yazar "Arıcan, Berk" ile LEE- Bilgisayar Mühendisliği-Yüksek Lisans'a göz atma
  • Öge
    Compressed domain image classification with sub-band data fusion
    (Graduate School, 2023-02-23) Arıcan, Berk ; Töreyin, Behçet Uğur ; Çarkacıoğlu, Levent ; 504181574 ; Computer Engineering
    Image compression algorithms aim to reduce the memory storage of an image without dropping image quality. There are two main approaches in compression methodology: lossless and lossy compression. Lossless compression focuses on preserving input image quality. Only redundant parts of an input image are removed with this approach. Another compression approach is lossy compression; since this method removes significant parts of the image permanently, the file size of an input shrunk with reduced image quality. JPEG2000 decoder/encoder algorithm based on Discrete Wavelet Transform (DWT) is developed to provide lossless and lossy compression. It also provides high scalability and accessibility in the compression stage. So, the images can be stored with any quality level. Images that require large amounts of memory space can store with JPEG2000. Processing vast archives of images compressed with JPEG2000 requires the entire decompression process followed by a highly computationally demanding image analysis process. The image analysis phase typically consists of machine learning applications, such as Deep Convolutional Neural Network (DCNN) models. Image classification is among the most common image analysis task. This study proposes a sub-band image-based classification method for JPEG2000-compressed images. The proposed work reduces memory usage and decompression time by using intermediate coefficients of JPEG2000 compressed images for the classification task. To that end, sub-band image coefficients of the Low-Low (LL), Low-High (LH), High-Low (HL), and High-High (HH) sub-bands are utilized. These coefficients are stored in the compression stage and can be accessed via partial decompression of the stored data without requiring total decompression. Each sub-band has unique details of an input image. LL sub-band images only consist of low-frequency details of an original image. Other sub-band images (LH, HL, HH) are deprived of low-band information. LH sub-bands store vertical, HL store horizontal, and HH store diagonal high-frequency information. In the scope of this work, we combined first-level LL, LH, HL, and HH sub-bands to represent both features with a single hybrid sub-band and fed to a DCNN for the image classification task. We proposed different sub-band fusion methods. The hybrid average sub-band is extracted as a first method by summing the LL sub-band with the average of high-frequency sub-bands. Thus, equally weighted high details are projected to the LL sub-band. In the second method, which considers high-frequency weight values, the maximum acquisition is applied to the high bands instead of averaging. The output of this process is added to the LL sub-band to create a hybrid maximum sub-band. Consequently, each proposed hybrid sub-band composition technique increases the high-frequency detail of the LL sub-band. We used two Remote Sensing Scene Classification (RSSC) datasets for the experiments; NWPU-RESISC45 and AID. NWPU-RESISC45 composes 31,500 scene images of 256x256x3 from 45 scene classes, and AID has 10,000 scene images with 600x600x3 size and 30 different classes. Since the images in these datasets are not compressed with JPEG2000, all images are compressed with JPEG2000 as a pre-processing step. We utilized DenseNet-121 to classify the data set. Our hybrid sub-band techniques improved the classification accuracy of the LL sub-band by %~1.62 and %~2.19, with only %~2.89 and %~0.61 milliseconds of additional test for the experiments in the NWPU-RESISC45 dataset. In experiments on the AID dataset, the proposed methods have more accurate classification performance than the LL sub-band by %~0.49 and %~0.93, with %~7.74 and %~1.48 milliseconds of extra processing time per image. Results show that a DCNN model can perform the partial decompression method with higher accuracy using both suggested sub-band composition techniques. Our work indicates that both proposed hybrid sub-band fusion approaches boost the high-frequency details in the LL sub-bands, allowing more details to be included in the image and improving classification performance while taking advantage of the partial decompression method.