Research On Underwater Image Segmentation Algorithm Based On Deep Learning

Image Semantic Segmentation is a very important research direction in the field of computer vision,which combines the traditional image segmentation and target recognition tasks.Image semantic segmentation is to divide each pixel in an image into its corresponding category and finally get a prediction result containing semantic information.Traditional image semantic segmentation mainly includes thresholding-based image segmentation techniques,edge-based image segmentation techniques,region-based image segmentation techniques and specific theory-based image segmentation techniques.Traditional image segmentation techniques are difficult to meet the requirements of practical applications in terms of segmentation accuracy and segmentation efficiency,especially for image processing of real-time scenes.For underwater images,the lighting conditions are unstable,and the light is affected by absorption,scattering and reflection of underwater objects,resulting in uneven color and brightness of the images.Noise such as water waves,suspended particulate matter,background clutter,diverse target shapes and sizes,and targets that may be obscured or mixed in a complex background,all of these factors can cause some interference to the accuracy of image segmentation.In recent years,with the continuous development of deep learning,researchers have proposed various algorithmic models based on deep learning and achieved certain results,but the process of image segmentation through deep learning models puts high demands on the performance of computers,and how to perform image segmentation efficiently and accurately is an urgent problem to be solved.In this thesis based on deep learning techniques,we propose image segmentation methods based on UNet network and Deep Lab network,respectively,to achieve better segmentation performance on the basis of guaranteed model complexity.The main work of this thesis is as follows:(1)A method for underwater image segmentation using a three-branch attention module to improve the UNet structure is studied,and a TAUNet model is built to achieve crosschannel information interaction through the attention mechanism.The studied TAUNet model achieves better segmentation results with almost no increase in the number of parameters and computational effort.The effectiveness of the proposed method is verified on the SUIM underwater dataset and the VOC dataset.(2)An improved underwater image segmentation method for UNet networks,called SKUNet,is investigated.SKUNet uses jump connections to fuse multi-scale features,thus improving the perception ability and recognition accuracy of the network for targets.Also,SKUNet employs an improved RSKFF module for feature fusion of multi-scale features,which further improves the network segmentation performance.In addition,SKUNet’s encoder module is able to pass these fused features to the decoder module to achieve more accurate and efficient image segmentation.Experimental results show that the proposed method achieves better segmentation results on several datasets.(3)An improved DeepLab v3+ network using new branches is investigated,which is able to extract more useful features from low semantic information to improve the segmentation performance of the network.In addition,a method to optimize the new branching structure using channel attention is discussed.The simple fusion of low-semantic features and high-semantic features does not necessarily improve the performance of the network.In this thesis using channel attention to focus on the channels of the feature graph can focus on the effective part of the feature graph more effectively and better achieve the fusion of low-semantic features and high-semantic features.The validation was carried out on the VOC public dataset and the SUIM underwater dataset,and the model proposed in this thesis achieved more obvious results.The aim of this thesis is to improve the method of underwater image segmentation,for which we have used the UNet network structure and introduced an attention mechanism and an improved RSKFF module for Deep Lab v3+.This approach combines different technical tools in order to improve the accuracy and robustness of underwater image segmentation.Specifically,we combine the UNet network with the attention mechanism in order to improve the model’s focus on the region of interest.In addition,an improved RSKFF module is introduced to further enhance the expressiveness and feature extraction capability of the model.Finally,experiments are conducted on the underwater image dataset and the performance of the method is compared with other commonly used image segmentation methods.The experimental results show that the method in this thesis achieves better segmentation results in the underwater image segmentation task,which proves the effectiveness of the proposed method.In conclusion,the underwater image segmentation method studied in this thesis combines a variety of deep learning models and provides a new idea for underwater image segmentation research.