Research Of Single-Stage Object Detection Algorithms Based On Deep Learning

Object detection is one of the long-standing classic research directions in the fields of computer vision and image processing.Based on the presence or absence of a proposal generation stage,object detection algorithms can be classified into single-stage detection algorithms and two-stage detection algorithms.This paper focuses on addressing the challenges of low detection accuracy for small objects and sampl eallocation in single-stage object detection algorithms.By analyzing and investigating the two mainstream single-stage object detection algorithms,SSD and FCOS,we center our efforts around the network model structure and training strategy to experiment with and improve single-stage object detection algorithms.Furthermore,we design models and experiments tailored for the Dongbaishan Wildlife Detection Project.The relevant research work of this paper is as follows:(1)To address the problem of low accuracy in small object detection in single-stage object detection,this paper proposes the EL-SSD model based on the SSD model by modifying the model from two sub-problems,the lack of contextual semantics,and the detection localization box bias.By constructing a bidirectional weighted feature pyramid,the predictive feature map features are fused,and the corresponding feature maps are decoded to obtain the position information of the features.Finally,the channel weights are reassigned for the corresponding feature maps,effectively alleviating the problem of missing contextual semantics.By studying the relationship between classification confidence and localization quality,an additional localization quality prediction branch is added,and a cascade clustering of localization confidence and classification confidence is constructed to perform non-maximum suppression on candidate prediction boxes,improving the localization accuracy in the post-processing stage and alleviating the problem of detection localization box bias.(2)To address the sample imbalance problem in single-stage object detection,this paper optimizes the training strategy based on the FCOS model,focusing on the two sub-problems of imbalance in positive and negative sample allocation and sample category imbalance,and proposes the LB-FCOS model.By treating the sample label assignment problem as an optimal transport assignment problem,we construct a soft label metric for sample classification cost and localization cost consistency and propose a dynamic spatial center-weighted cost to replace manual prior center sampling.We then apply positive and negative soft weighting to the positive sample loss function,alleviating the unequal positive and negative sample issue during target detection training.By adding category balance weights to the loss,we improve the detection effect of rare categories in long-tailed data and alleviate the sample category imbalance problem during target detection training.(3)In response to the challenges of low detection accuracy for small target animals and long-tailed data in the Dongbaishan Wildlife Detection Project,this paper combines the research on small target detection and sample allocation,proposes the LBEL-FCOS model,and creates a dataset.By adding a coordinated bidirectional multi-scale feature fusion module to the anchor-free detector,we allocate samples using a soft label optimal transport assignment strategy and introduce category balance weights to the positive sample classification loss,improving the detection accuracy for rare animal species.On the custom wildlife dataset,the mean Average Precision of the LBEL-FCOS model reached 74.2%,an improvement of 4.6% compared to the baseline FCOS model.