Research On Small Object Detection Algorithm In Multi Object Scene Based On Deep Learning

With the advancement of computer vision technology,object detection has been considered a key technology.Object detection is an important task in the field of computer vision,and small object detection has always been a difficulty and challenge in the field of object detection due to its difficulty in extracting effective features.In order to improve the detection performance of small objects in multi-scale target scenarios,which includes both small and large objects,many studies have conducted in-depth exploration in network architecture,training strategies,data processing,and achieved significant results.In this thesis,we propose a new attention-mechanism-based bidirectional feature pyramid architecture named Ag BFPN,to enhance the semantic and spatial information transfer between each level of scale feature map.Ag BFPN uses the Channel AttentionDirected Fusion(CAGF)module and the Spatial Attention-Directed Fusion(SAGF)module to enhance feature fusion,CAGF mitigates the loss of information caused by reduced channels and better transmits semantic information from high to low level features.SAGF transfers rich spatial information from shallow features to deep features.Experiments show that Ag BFPN achieves higher detection accuracy of small targets and average detection accuracy of multi-scale objects.In this thesis,based on the proposed attention-mechanity-guided bidirectional feature pyramid architecture Ag BFPN,an anchor-less optimized detection head called CAFHead is designed to optimize small target detection by integrating cascade query and adaptive tag assignment.CAFHead introduces a cascaded sparse mechanism,that is,low-resolution feature maps are used to initially locate the approximate location of small targets,and then the low resolution feature maps are used to guide the calculation on the high-resolution feature maps.In other words,it does not calculate all the spatial positions of the feature map,but only uses the detection head to detect the positions where small targets may exist,thus saving the calculation cost.In addition,CAFHead incorporates a label allocation strategy that adaptively assigns the most appropriate positive and negative samples for each object,directly predicting objects at each location without the artificial design of Anchor.There is no need to set the size and proportion of anchor points in advance,nor need to use manual screening or clustering method to determine the adaptive label allocation strategy of anchor points,so small targets can be detected more accurately.Experiments show that the addition of CAFHead achieves a higher average precision of small object detection.