Research On Object Detection Algorithm And Application Based On Visual Saliency

The human visual system can acquire and process information autonomously and efficiently in the unstructured visual signals in a large number of complex scenes,the process is governed by a visual attention mechanism and this mechanism can help people choose visual information quickly.In order to simulate the human visual attention mechanism,many experts have proposed a series of visual saliency detection models in the fields of computer vision,image processing and physiology.Due to differences in feature selection,saliency calculation and others reasons,the detection model is difficult to meet the actual needs of computer vision tasks.How to combine the knowledge of various disciplines to design a saliency detection model that conforms to the human attention mechanism and meets the practical application is worthy of further study.This thesis will study the saliency object detection from two aspects:traditional methods and deep learning methods.The main contents are as follows:For problem of difficult detection with target contact boundary and multi-target in the original MBD detection model,this thesis designs an MBD saliency object detection model based on visual diffusion,and the model realizes the diffusion of images from outside to inside and from inside to outside.Finally,the saliency map obtained by combining the two update methods.This thesis uses the classic detection database and corresponding evaluation indicators.From quantitative and qualitative results analysis,this thesis effectively solves the various problems of traditional MBD and has high accuracy and other advantages.The traditional method is not as good as the current saliency detection model based on deep learning on the detection ability of common object modes,however many deep learning detection models are relatively simple in the combination of feature layers,and multi-scale processing capabilities are flawed.For these reasons,this thesis designs a Multi-scale Feature Sharing(MSFS)saliency object detection network.The MSFS network consists of a basic VGG16 feature extraction module and a multi-scale feature sharing connection module.The network fuses different levels of convolution information through a shared connection during the deconvolution generation process.Compared with traditional detection methods and existing partial deep learning methods,the proposed network structure is novel,the scale is small,the accuracy is high,and the effect is satisfactory.The two visual saliency object detection methods designed in this thesis have completed the image saliency object detection task well and achieved satisfactory results compared with the existing methods.Among them,the diffusion-based MBD saliency detection model is suitable for the case where the amount of data is insufficient and the scene changes frequently,and the MSFS network is suitable for the case where the data volume is sufficient and certain types of targets are detected.In this thesis,different detection methods are provided for different application scenarios,and verification and comparison are carried out in practical application cases.