Underwater Single-pixel Imaging Method Based on Object Search and Detail Enhancement
In the realm of underwater imaging,current Single-Pixel Imaging(SPI)technologies are grappling with a substantial challenge when deployed in intricate optical environments.Predominantly,conventional methods focus on reconstructing the overall representation of the object,which inherently restricts their ability to optimally restore and emphasize the minute details within the image.This inherent limitation has a profound impact on the overall quality and resolution of the reconstructed images,particularly in scenarios where precise analysis and interpretation require high levels of fidelity.In light of this pressing issue,we propose a methodological approach for underwater single-pixel imaging that ingeniously integrates two pivotal mechanisms:object search and detail enhancement.The core essence of the proposed method is bifurcated into two main objectives.Initially,the object search component deploys intelligent algorithms that meticulously analyze the fluctuations in pixel intensities across rows and columns within the reconstructed image.So it expertly discriminates between the object area and its surrounding background,effectively singling out and amplifying the object signal while simultaneously attenuating background noise.This strategic isolation significantly enhances the contrast and visual prominence of the targeted object.On the other hand,the detail enhancement facet of our methodology harnesses state-of-the-art machine learning techniques,specifically leveraging a part-based model embedded within a Convolutional Neural Network(CNN)architecture.This sophisticated model specializes in discerning and learning the complex,fine-grained features encapsulated within the collected light intensity data.Upon extracting these learned attributes,the methodology proceeds to refine and accentuate the detailed aspects of the object within the reconstructed image,thereby elevating its overall resolution and sharpness.To rigorously substantiate the dependability and efficacy of our novel technique,we have conducted an extensive series of experiments in both space and underwater settings.During the preliminary experimental phase,we concentrated on five distinct alphabetical objects-"I","O","P","E",and"N"-comparing the performance of our method against the Traditional Ghost Imaging(TGI)and Different Ghost Imaging(DGI)methodologies.We carried out meticulous measurements of the Contrast-to-Noise Ratio(CNR)and spatial resolution of the reconstructed images,as well as closely examining grayscale values at specific points such as the slits within the letters"O"and"P".The experimental results underlined that,under space conditions,the proposed method surpasses conventional approaches by successfully maintaining and enhancing the intricate detail information of the object,thus leading to a significant improvement in the reconstructed image quality.Additionally,to prove the robustness of our method across various sampling rates,supplementary tests were performed in the space environment.By calculating the CNR and resolution of reconstructed images at different iteration counts,we empirically demonstrated that even at lower sampling rates,our method consistently delivers enhanced detail,showcasing its adaptability and versatility.Taking our experimentation one step further,we ventured into highly turbulent water conditions,executing over 1 500 iterations of transmission and reflection SPI experiments.Despite the challenging nature of these dynamic and unpredictable environmental conditions,the proposed method exhibited superior performance,solidifying its reputation for resilience and reliability under diverse circumstances.The comprehensive experimental findings provide compelling evidence of the merit and value of our innovative underwater single-pixel imaging method.It decisively demonstrates that,whether the imaging context involves unknown space environments or intricate underwater landscapes,our method can reliably and accurately reconstruct high-quality object images,even with limited sampling rates.
NETL
NSTL
万方数据
