Semantic focus fusion based on deep learning for deblurring effect / Ismail

Faculty of Electrical and Electronics Engineering

Thesis (Doctor of Philosophy) -- Universiti Malaysia Pahang – 2024

Includes bibliographical references

High-quality images are essential in image processing as they can provide accurate image information for both humans and machines. They have been used in many industrial fields, such as medical diagnostics, robotics, and surveillance. Due to the limited depth of field (DOF) of camera lens, the camera cannot generate high-quality images without blurred region images. In a rapid development of intelligent computation, such as deep learning algorithm, multi-focus image fusion methods indirectly being involved, such as CNN and PCA Net architectures. Nevertheless, they still lack of accuracy and stability. The deficiencies are affected by the shortage of CNN and PCA Net methods to establish accurate distance pixels. The new method utilizes another pixel property to extract a focused image It is built upon the pixel density method and classifies image pixels according to suitable classes. The method is termed semantic focus fusion for deblurring effect. It employs deep learning architecture to extract focus and blurred features. It projects pixels into focus or blur classes. Semantic focus fusion architecture contains focus extraction, feed-forward mapping, and upsampling modules. The focus extraction module comprises a deep learning model and a pyramid filter. The feed-forward mapping module is a mapping module connecting the beginning layer to the last layer of the network. It maintains the original information of the input image and increases the accuracy of the classification. On the other hand, the upsampling module increases the resolution of the output image to the size as the input image. Precision, F1-score, and accuracy indexes evaluate the correctness of the predicted focus map with 0.995739, 0.665717, and 0.66531 for F1-score, and accuracy of semantic focus fusion. It gets index scores higher than CNN with precision, F1-score, and accuracy are 0.66949, 0.572463, and 0.551008, respectively. At the same time, PCA Net has 0.562647, 0.529477, and 0.437959 of precision, F1-score, and accuracy successively. The weight performance, SSIM, and PSNR indexes of semantic focus fusion are 7.698, 1.00, and 83.84 dB. It can be said that those indexes are also higher than both CNN and PCA Net. Meanwhile, CNN has 7.021, 0.9183, and 28.28 indexes, and PCA Net has 3.388, 1.00, and 76.21 indexes. The semantic focus fusion classifies focus and blurred pixels more accurately than CNN and PCA Net methods based on its capability to deblur a blurred effect in fused images derived from its high reliability. Finally, the semantic focus fusion with the feed-forward mapping model has enough ability to increase image quality in many industrial applications.