Multi-Scale Object Detection Based On Deep Learning

The rapid development of deep learning has boosted the popularity of CNN-based object de-tectors.However,multi-scale object detection,especially small object detection and precise object localization,remains an unsolved challenge.The difficulty of small object detection stems from little related information and feature coupling with other objects,while the difficulty of localiza-tion stems from improper object representation containing redundant background.To solve these problems,this thesis proposes corresponding solutions and a stronger multi-scale object detector.The main contributions can be summarized as:1.This thesis proposes an extended feature pyramid network(EFPN)to enhance small ob-ject detection.EFPN introduces an extra pyramid level to decouple features where more regional details are encoded for small objects.A novel feature texture transfer module ef-ficiently captures more regional details for the extended pyramid through feature-level SR.Moreover,this thesis introduces a cross resolution distillation mechanism to transfer the ability of perceiving details across the scales of the network.The experiments suggest that EFPN improves the generalized performance of small object detection more efficiently.2.This thesis proposes the detection head based on Edge Representative Point Set for precise object localization.Edge Representative Point Set uses prior edge extreme points to locate objects.The detection head aligns features with object area by way of stage design and region feature extraction.Besides,the proposed edge feature sampling utilizes location information from edge representative point for the position-sensitive localization branch.The experiments suggest that the detection head based on Edge Representative Point Set locates objects better via edge extremes.3.This thesis proposes a multi-scale edge-sensitive detector(MSESDet)that is based on EFPN and Edge Representative Point Set.In addition,MSESDet complements the whole multi-scale detection pipeline,including the framework,label assignment,training scheme and postprocessing.MSESDet is both state-of-the-art on public datasets and practical in real applications.