Research On Detection Methods Of Ship Target Image Under Visual Narrowing

In recent years,target detection technology for satellite remote sensing images and camera video images has developed rapidly,especially in the civil and military fields.Scholars in various countries have made certain development and breakthroughs in ship target detection.However,in the research’s process,the ideal experimental environment is far away from the real one.The interference of noise,complex background and other factors will affect the final detection effect.However,the existing theoretical results,such as the detection rate,detection speed and algorithm versatility still have shortcomings.Therefore,It is the focus and key of this paper to effectively detect ship targets in complex environment.With the rapid development of machine vision,detection technology has been widely used in all walks of life,because it has the characteristics of non-contact,real-time,reliability,high automation and high precision.Narrowing refers to the process in which people concentrate their cognition,thinking,feeling or emotion in a certain direction,and the scope involved is becoming more and more limited and narrowed.They choose one attribute from multiple attributes or narrow one attribute.In this paper,a semantic algorithm of scene narrowing and topic narrowing is added to Faster R-CNN convolution neural network,which accelerates the detection speed.In addition,the idea of sub-network and main network narrowing exists under certain circumstances.Therefore,narrowing is a new concept for image processing in machine vision,also has a certain positive significance.As we all know,the process of target detection includes:Preprocessing,image segmentation,feature extraction and classifier detection.In view of the research focus of this paper:image feature extraction and neural network detection,the following results have been achieved:1.In order to improve ship target detection in high-resolution remote sensing images,the paper proposes a ship target detection method based on SRM segmentation and feature extraction of hierarchical line segments.Because high resolution remote sensing image is a large-scale image with rich details and complex texture,it is difficult to analyze.Therefore,firstly,reducing the image size by downsampling method.Then,according to the line segment extraction method,the regional line segments are formed by merging the adjacent line segments under the threshold with searching and updating the layered line segments.Finally,the fast detection of ship targets is realized by comparing with the ship’s length and width characteristics.The experimental results show that the proposed method can effectively detect ship targets in high resolution remote sensing images,and the performance of the hierarchical line segment feature extraction method is better than other commonly used methods.At the same time,the ship detection method in this paper does not need too many parameters and training samples.2.In order to solve the problem of high false detection rate from video images~’ship target detection which are based on hierarchical line segment feature extraction,the paper presents a ship target detection method based on Faster R-CNN convolution neural network with narrowing semantics.Firstly,the frame images which are extracted from the camera video are labeled.Secondly,the scene is narrowed to the ship target region.Then the narrowed images are trained in the built deep convolution network with topic narrowing function.Finally,the ship target detection is completed according to the four basic target detection steps:region proposal network,region of interest pooling,classification and non-maximum suppression.The experimental results show that this method is suitable for remote sensing images and video images,and can effectively detect ship targets in images.The detection time of Faster R-CNN convolution neural network with narrowing semantics is less than that without narrowing semantics.