An Underwater Sensing System Based on Color Correction and Depth Information Dehazing
Mao Zhaoyong,Liu Nan,Chen Gangqi,Hou Dongdong,Shen Junge
DOI: https://doi.org/10.3788/gzxb20245306.0610001
IF: 0.6
2024-01-01
ACTA PHOTONICA SINICA
Abstract:The core of an Autonomous Underwater Vehicle(AUV)lies in its ability to accurately perceive objects and the surrounding environment.With advancements in underwater optical vision sensor technology,optical imaging for environment perception is now feasible.Despite progress in object detection,underwater images'inherent degradation poses challenges.High underwater pressure complicates distance information acquisition,leading to limited training datasets.Moreover,the degradation and blurriness of underwater images often obscure object features.To enhance AUVs'capabilities in distance perception and scene reconstruction,research is increasingly focusing on precise localization and depth scene construction in underwater scenarios.To this end,this paper introduces an underwater visual perception system which incorporates color correction and depth information dehazing to overcome these challenges.Specifically,we propose an improved color correction method that combines white balance and adaptive histogram equalization for effective white balance and histogram adjustments to original images.This approach effectively mitigates the common issue of red artifacts in underwater images,thus rendering the images more realistic.Additionally,our method leverages white balance adjustments to enhance overall image contrast,thereby improving feature clarity.Moreover,to address the challenge of data insufficiency in underwater distance perception tasks,we have developed an improved fusion enhancement method.Through this approach,we establish an underwater monocular image dataset.Specifically,we collected a large number of underwater images from the Internet and enhanced them using the aforementioned image enhancement method.Building upon this,we integrated a monocular depth estimation network into our framework,where the depth estimation network is trained on the collected underwater images in an unsupervised manner.This approach provides depth map information,which is essential for subsequent image dehazing within the framework.Furthermore,to address the mis-detection issue in object detection caused by image degradation,we developed a novel underwater dehazing method.Note that the depth information generated by the monocular depth estimation network provides a more accurate modeling than prior knowledge,thus further enhance the dehazing performance.This method not only enhances image quality but also effectively clarifies degraded and blurry images,when it is incorporated into the proposed underwater imaging perception framework.To achieve more precise object localization,we propose a novel channel reordering network based on center point detection.This method effectively incorporates fine-grained features from the shallower layers of the convolutional neural network into the deeper layers.It should be noted that this anchor-free method effectively enhances feature extraction for small and dense objects.The efficacy of this method was demonstrated through extensive experiments on multiple datasets,including recovery experiments on underwater images.Extensive experiments were conducted to validate the method's ability to restore true terrestrial colors and to accurately perceive relative distances in underwater scenes.Additional experiments validated the method's high-precision object perception capabilities both within and across domains,achieving high performance levels on the URPC-Color and the URPC-Dehaze datasets.Furthermore,a comparison was made with various advanced one-stage models on the URPC dataset.Our method achieves an in-domain object detection accuracy of 78.2%,representing a 4.6%improvement over the baseline CenterNet.Moreover,category-wise accuracy performance shows that our method surpasses all other methods by a large margin,further indicating its effectiveness in underwater scenarios.In cross-domain detection experiments,our method achieves competitive results with an 81.5%object detection accuracy on the UTTS dataset.This further indicates the cross-domain capabilities of our method in underwater scenarios.The color correction and dehazing experiments highlighted the method's ability to enhance image quality and more effectively perceive scene depth and object information.