Research And Implementation Of Small Target Detection Method In Images Based On Deep Learnin

Target detection,a key area of computer vision in recent years,has been given much consideration both theoretically and practically.Small targets are ubiquitous in many scenes,thus making the study of small target detection of great practical importance and applicability in image target location and recognition.The small target has low resolution,less information,small area in the image and variable aspect ratio,which leads to the problems such as less effective features extracted by the algorithm,large background interference and difficult to locate accurately.The current target detection techniques are far from perfect,and it is effortless to lead to erroneous or missed detection of minor targets.In this paper,an improved small target detection model AU-DC-YOLOv5 based on YOLOv5 is proposed,which can effectively solve the above problems.Optimizing and altering the semantic details of minuscule targets significantly enhances the precision of their detection.The research of this paper focuses on two aspects:(1)This paper optimizes YOLOv5 feature extraction by utilizing deformable convolution and cross convolution,constructing the DC-YOLOv5 model to address the issue of inadequate feature extraction by traditional convolution.To begin with,deformable convolution is substituted for the Conv module of the shallow network,and the receptive field is adjusted automatically to more effectively concentrate on small targets.Secondly,by introducing cross convolution,the C3 module of shallow network is replaced by C3 Cross module,and the edge information is extracted by vertical and horizontal gradient information in parallel,so as to obtain clearer edge features.at the same time,more similar feature information is retained through cross-graph convolution between different feature images.Experimental findings demonstrate that this technique can draw out more adequate semantic data and enhance the network’s functioning.(2)This paper optimizes the YOLOv5 model,utilizing feature enhancement and attention mechanism to facilitate feature fusion,and builds the AU-DC-YOLOv5 model on the foundation of the DC-YOLOv5 model initially established.First of all,an up-sampling feature enhancement module is constructed based on inflated convolution to expand the receptive field of the feature graph,fully extract context information,and enhance the feature representation in the process of transmission.Secondly,the global maximum pool is added to the ECA module to improve the ECA.The attention mechanism module ECA-SAM,proposed in light of the enhanced ECA and CBAM modules,is beneficial in optimizing feature fusion mode,extracting essential data,and enhancing the model’s focus on small targets,thus augmenting the algorithm’s detection accuracy.Finally,the loss function of the new metric based on Io U and normalized Wasserstein distance is designed,which is more suitable for small target detection scenarios.The experimental results show that the model enhances the feature representation in the process of transmission,performs better feature fusion,and effectively improves the ability of small object detection.