Research On Anchor-free Detector Based On Feature Enhancement And Center Sampling

Object detection is a hot topic in the field of computer vision.It is not only widely used in practical scenarios such as industrial detection and auto-driving,but also plays an important role in subsequent more complex visual problems such as face detection and instance segmentation.With the development of deep learning,deep neural networks have been widely used in object detection due to their powerful feature extraction and excellent expression performance,breeding many classic detectors and showing great application value in industrial field.Although the object detection algorithm has achieved great success,there still exist huge problems in different aspects such as scale variance and sampling: first,the change of scales requires good feature fusion methods.However,the common feature fusion methods lack the choice of features,the fusion method is relatively simple,and the information transmission between different levels of features is difficult and has certain limitations.Second,anchor-free detectors has become a new research hotspot in recent years.However,due to the immature development,there are still drawbacks such as poor sample in the edge of objects and loss functions that cannot be differentiated to deal with certain scenarios.These problems limit the further improvement of anchor-free detectors.In order to solve the above problems,this article proposes a interconnected feature pyramid network based on attention mechanism for feature enhancement,and improves on multitasking,loss function design and sampling strategy,and then gets an excellent anchor-free detector.The main work of the detector is as follows:1.This paper proposes a interconnected feature pyramid network based on attention mechanism for feature enhancement,which achieves good feature selection and allow different levels of features to be freely fused.At the same time,it works on the anchor-based detectors and anchor-free detectors,with good migration.2.A object detection algorithm C-FCOS based on center sampling is proposed.It uses a more accurate composite confidence to guide the non-maximum suppression process.At the same time,it uses a improved regression loss function based on shrink initialization and dynamic weights,and uses a center sampling strategy based on edge elimination to improve the quality of samples.3.In this paper,the model input is optimized,and the acceleration library is used for data preprocessing and inference,thereby greatly improving the speed of the detector and increasing the practicality of the algorithm.This article involves anchor-based detector,including multi-stage and single-stage detectors,as well as the latest anchor-free detectors.This article conducts a large number of experiments on COCO and the result shows the effectiveness of the IFPN and the good performance of C-FCOS.On this basis,the transplantation improve the efficiency of C-FCOS,and has wide application value.