Research On Multi-Level Salient Perceptual Object Detection And Task-Driven Model Application Study

The goal of salient object detection is to accurately locate and segment the objects or regions that attract the most human visual attention by analyzing the content of the image.As an important pre-processing technique,salient object detection does not depend on the semantic class of the object to be detected and does not limit the complexity of the scene,and has been widely used in several types of vision tasks to help them efficiently capture the most relevant regional features in the task.Currently,deep fully convolutional neural networks have dominated the field of salient object detection due to their powerful multi-scale feature extraction capability.However,existing salient object detection algorithms usually suffer from incomplete extraction and low contour resolution for accurate detection in non-connected domains,and are often limited by the loss of task-relevant features when deployed to practical application scenarios,and poor generalization to meet the prediction needs of multi-target tasks in different realistic scenarios.Therefore,a more advanced object detection technology is an urgent issue that needs to be addressed to achieve excellent performance with a higher resolution,better detection rate,more efficient,and in a more robust manner.A novel perceptual matching-based correction algorithm model is proposed to address the problems of incomplete and inexact detection of non-connected regions/objects for salient object detection,low resolution of contours,etc.Firstly,multi-scale multi-level complementary perceptual contents are extracted and discriminative feature details of salient objects are localized in the local-level regions of foreground,contours and background using multi-connection channel and spatial attention mechanisms on channel and space in the global level,image-level attention is used to detect multi-significant objects,non-connected regions,etc.Then,based on the separated perceptual contents,forward additions and reverse deletions are differentially performed to correct error-prone regions and further maximize the separation of multi-level features in holes,non-connected regions.Locally accurate discriminative features and globally complete multi-objective salient objects are located by bi-directional aggregation to segment and correct the attention for feature fusion,maximizing the extraction of mutual information knowledge between layers and neighborhoods.Based on top-down guidance with bottom-up feature mapping,a subjective structure loss objective function is proposed to minimize the foreground,contour and fore/background errors of saliency map and ground truth maps to find the global optimal solution.The optimal performance is achieved by comparing the analysis with 12 advanced salient object detection methods on five benchmark datasets.Moreover,it runs fast at 27 fps in real time when processing 256×256 images,which is beneficial for practical applications.Salient object detection in vision tasks can assist applications to detect the most relevant targets of the task and improve the computational efficiency of massive data.Facing the demand for complete and fine prediction of task features in task-specific applications,a knowledge transfer-based model for adapting salient object detection to complex application scenarios is proposed,which transfers generic knowledge transfer to the student module through the teacher module,so that it has a generic knowledge reserve of salient objects to learn task-specific salient knowledge and extract task-driven features.In the application,safety supervision,as an important task in smart oilfield,faces difficulties such as wide operation scope and difficult management.Therefore,Oilfield-Helmet,the first salient object detection dataset for oilfield safety supervision task,is proposed based on the site of Qijia depression work area in Daqing oilfield to detect whether workers wear helmets correctly to ensure industrial safety.Compared with the state-of-the-art general-purpose model on Oilfield-Helmet dataset,it is able to learn task-driven feature knowledge with higher check-all rate,higher accuracy rate and smaller error.The overall performance of the task-based model is better for qualitative and quantitative analysis on real-world scenarios-360° omnidirectional images,autonomous driving images,and industrial robot images.Further,the low performance of the generic model on Oilfield-Helmet compared to the high performance data on the task-based model validates the effectiveness and high challenge of the oilfield safety surveillance dataset compared to other task scenario datasets.