Object Detection Based On Deep Saliency Analysis

The task of object detection is to identify the object from images quickly and accurately,the object detection has been developed rapidly with the rapid development of artificial intelligence and computer vision in the field of image processing.However,with the rapid popularization of various imaging instruments,the image data resources are becoming huge,so it is urgent and essential to deal the image data quickly and efficiently with the help of computer.In order to characterize the image,we have to find and extract the most valuable information from the complicated image data as soon as possible.The main task of saliency object detection is that the computer utilizes various characteristic information of object in the image to simulate the human visual system and understand the image effectively to extracte the areas of which human concern in the scene,so we can effectively liberate the human and financial resources.It is a very important research direction in the computer vision and machine learning fields.In this context,in view of the existing problems of object detection based on the saliency analysis,we focused on the simulation of human visual attention mechanism,aiming at the uniform prominence of salient region and suppressing the redundancy background.Then,based on this,we studied and discussed the saliecy analysis and also applied to object detection on remote sensing images,the main work summary is as follows:1.We proposed a method of image saliency analysis based on Markov absorption model,this method characterize pixels in the image as a node of a multilayer undigraph model.In order to optimize the edge weights,the edge of the weighted adjustment graph model based on spatial relationship between layer and layer is proposed,and the background prior knowledge is also introduced into the model to enhance the contrast between foreground and background.In order to detect the characteristic of node more accurately,we design and use the probability of being absorbed to characterize the object more effectively.The detection result of algorithm is closer to the real ground truth,which can better highlight the object area,and restrain the background area.It can keep the boundary of the object well and avoid the problem of over segmentation,which improve the detection rate meanwhile.2.We proposed a saliency analysis algorithm based on deep iteration scheme,which is based on the superpixels segmentation of previous proposed algorithm.The method uses global structural loss supervised information between the salient image and ground truth to adjust the salient feature map of the next scale automatically.Until the loss information is less than the certain threshold,the object detection result can be obtained through deep saliency analysis,and we find that the method has successfully detected the object on the complex optical dataset.3.On this foundation,we also studied the object detection in remote sensing image.Because the remote sensing image always with large scene and small relative target,so the detection result will be easily missed by the number of hyper-pixel segmentation,and the detection precision is decreased.Therefore,we proposed the object detection method of remote sensing image based on spatial correlation filtering and deep significance analysis.It can quickly locate the significant target in the image,and remove the background area unrelated to the detection task,which has essential practical application value.