Research On Technologies Of Image Object Detection Based On Contextual Information

Image object detection technique is one of the fundamental research problems in Computer Vision field.The main task is localizing and classify specific objects in images.Most of the modern target detection techniques are based on deep learning methods,leveraging the powerful image feature learning capabilities of neural network structures to achieve object region category prediction and bounding box position regression.However,such methods only use visual feature information and process each object independently,lacking the analysis of the intrinsic relationship between the scene specific information and the object,thus the overall detection accuracy is limited.And it is also difficult to adapt to the demand for accurate Recognition of multiple objects and in complex scenes.In this paper,we study how to establish an effective image scene context analysis model to improve the Recognition performance of object detection model.For the contextual information in images,the main research methods and work in this paper are as follows.(1)A multi-scale feature context-based object detection method is proposed.There is often a certain visual correlation between visual objects in real scene images and their neighbors.In order to preserve the continuity of multiscale features in the scale space and to constrain the noise information in the feature fusion process,this paper proposes an improved method of multiscale feature fusion that uses the feature context to improve the accuracy of the network.(2)A spatial context-based object detection method is proposed.Many of the relevant objects in a real scene will appear as specific spatial relative positions.For example,"people" and "bicycle" usually appear in the form of above and below position.In order to analyze and utilize the contextual information of spatial location,this paper proposes a spatial relationship inference method based on deep neural network structure,to quantitatively learn and represent the spatial relevance between objects,so as to improve the Recognition accuracy of the object detection network.(3)An object detection model based on multi-scale feature context and spatial context is proposed,jointly using two types of contextual information.Since the information sources of multi-scale feature context and spatial context are independent of each other,we design a complete network structure to apply two types of contextual feature information for joint prediction.The model is based on a generic two-stage target detection structure with a multiscale feature context and spatial context approach.Experiment results show that the overall accuracy of the detection model is improved by the effective use of multi-scale feature context and spatial context information,especially for hard targets in complex scenes.