Research On Underwater Image Enhancement And Object Detection Based On Transfer Learning

Underwater target detection aims to locate and identify targets in underwater scenarios,which is of great significance in underwater applications such as ocean detection and monitoring,and autonomous underwater vehicles.However,images obtained in complex underwater environments often have serious degradation,which affects the implementation of high-level visual tasks such as underwater target detection.The underwater image enhancement algorithm can improve the image degradation and improve the quality of underwater images.But there is a lack of underwater ground-truth images,the generalization performance of the model trained by the learning method on the synthetic data is limited.And since the enhancement and detection tasks have different optimization goals,when only considering the enhancement effect of the enhancement algorithm,the enhanced image may not really be helpful to the execution of the target detection task.How to effectively enhance underwater images and improve the performance of underwater detection target algorithms has become a hot and difficult issue in underwater computational vision.Aiming at the problem that underwater image degradation in real scenes affects the execution of high-level tasks such as underwater target detection,this paper first proposes a two-step domain adaptation underwater image enhancement method,which provides a new idea of using transfer learning to study underwater image enhancement.Further,in order to make the enhanced image better applied to practice,an underwater image enhancement method based on deep transfer learning and color restoration model is proposed on the basis of the first work.Finally,a joint underwater image enhancement and target detection algorithm in real-world scenarios is proposed.The underwater image enhancement and target detection are optimized simultaneously,making the underwater image enhancement algorithm can really improve the performance of target detection task.The main innovations of this paper are as follows:First,due to the lack of paired training images underwater,many methods train models on synthetic underwater images,resulting in limited generalization performance of such methods when enhancing real-world underwater images.To solve this problem,inspired by transfer learning,a novel two-step domain adaptation method for underwater image enhancement is proposed,which transfers image dehazing to underwater image enhancement.Through domain adaptation for style transfer and domain adaptation for image enhancement,the cross-domain transfer from underwater domain to air domain is realized.And this method does not require training on synthetic underwater images and does not rely on ground-truth underwater images.Second,to counter the problem of low visibility,blurred details and color distortion in the images obtained by the vision system when underwater vehicles detect the ocean,based on the first work,an underwater image enhancement method based on deep transfer learning and color recovery model is proposed.The color restoration model is embedded in the domain transformation module to restore a more natural underwater image that meets the physical model.The coarse-grained image similarity calculation is added before the domain transformation to improve the performance of the algorithm,effectively remove the color difference in the image,improve the quality of underwater images,and help the execution of underwater detection tasks.Third,the degradation of underwater images affects the detection accuracy of underwater target detection tasks.Many underwater image enhancement algorithms simply pursue the effect of image enhancement without considering the impact of enhanced images on subsequent underwater target detection tasks,resulting in different optimization schemes for image enhancement and object detection.In view of the above problems,a joint underwater image enhancement and target detection method in real-world scenario is proposed,which optimizes image enhancement and target detection at the same time,and generates underwater images conducive to underwater target detection while improving underwater image degradation,thus improving the accuracy of underwater target detection.