Research On Small Target Detection Method Based On Multi-scale Feature Fusio

As artificial intelligence technology advances and evolves,computer vision technology has also stepped onto the stage.As one of the fields,object detection is broadly applied in many tasks such as industrial defect detection,intelligent traffic,smart medical care,and intelligent detection.Nowadays,deep learning based on object detection models have taken to a new level,achieving satisfactory results in detecting large and medium-sized targets in images.However,the detection accuracy of small targets is still far inferior to that of large and medium-sized targets.Therefore,small target detection still has high research value in both academia and industry.This article selects a single stage detection model with both speed and accuracy,aiming to improve the accuracy of small object detection.Two small object detection algorithms based on multi-scale feature fusion are designed for general and dense scenarios.The main work of this article is as follows:1.In response to the problem of low accuracy of small targets in general scenarios,an SSD model suitable for general scenarios was selected as the benchmark model,and a small target detection model integrating multi-scale features and mixed attention was proposed.The model designed two multi-scale feature fusion paths.Among them,the bottom-up path conveys spatial details to the deep layer,and the top-down path conveys Semantic information to the shallow layer.Finally,the hybrid attention mechanism is combined to enhance the internal correlation of features and reduce the loss of details and Semantic information in the convolution process.Experimental results show that the proposed algorithm outperforms the SSD model when the input image size is 512 ×512,a m AP of 84.6% was achieved on the PASCAL VOC dataset and 89.6% on the HRRSD dataset,indicating good generalization of the proposed model.2.In order to solve the problem of poor detection accuracy of YOLOv5 detector for small targets in dense scenes,a small target detection model integrating Swin Transformer and multi-scale feature decoupling is constructed.This model embeds the Swin Transformer structure into the CSP module of the backbone network,which enables the model to focus on the features that are useful for the current task and enhance the global information of the feature map on the premise of increasing less computational effort;Then,two multi-scale feature fusion modules were designed in the Neck section to decouple the classification task from the regression task,reducing the conflicting information brought by the regression task and the classification task.Experiments were conducted on the Vis Drone dataset,and the experimental results showed that compared with YOLOv5,the proposed algorithm improved by 4.53% on AP50 and 5.67% on AP75,proving that the algorithm can effectively improve the detection performance of small targets in dense scenes.