Research On Small Object Detection Algorithm Based On YOLO Network

In recent years,deep learning-based object detection techniques have achieved excellent results in routine tasks.However,in small object detection,the technology still faces enormous challenges and constant breakthroughs.The main difficulty of small object detection is that it has few pixels and a small proportion in the image,which makes it easy for large-scale convolutional neural networks to lose feature information during feature extraction,and the background information in natural scenes interferes with small objects to a great extent.Therefore,how to fully extract feature information of small-sized objects and accurately obtain their semantic and positional information in the image has become the core of solving the small object detection task.Based on the above problem,this article conducts research on small object detection based on the YOLOX algorithm.The specific work is as follows:(1)In response to the problem of interference from complex background information in feature maps,this article proposes a new anchor-free small object detection algorithm,BiYOLOX.First,the attention mechanism in the spatial and channel mixing domain is introduced to strengthen the extraction network’s focus on beneficial detail semantic information.Then,multiple multi-scale skip residual connection paths are added to the feature fusion network to bring rich feature information to different stages of the feature map and reduce the loss of feature information for small-sized objects.Finally,the detection head is modified for smaller object sizes by discarding the deepest detection head and adding a shallow detection head with a size of 160×160.Experimental verification on the Visdrone and DOTA datasets shows that the m AP of the Bi-YOLOX algorithm is 40.66% and 84.46%,respectively,which is 7.56% and 8.48% higher than the baseline network.Furthermore,the model’s effectiveness is confirmed by a reduction of 4.74 MB in network parameters.(2)To address the issue of small objects having less inherent feature information and easily losing semantic details,this paper proposes an improved algorithm called MMF-YOLO based on the Bi-YOLOX algorithm.Firstly,in order to reduce the feature redundancy brought about by massive fusion,an adaptive fusion factor alpha is added to each path,and the weights of different paths are dynamically updated.Secondly,in order to reduce the loss of initial feature information in the spatial dimension,the internal structure of CBAM is modified in this paper.The initial features are weighted by both CAM and SAM,and the weighted results of both are added and outputted,avoiding the partial loss of initial feature information in both channel and spatial dimensions.Finally,experimental verification is conducted on the Visdrone and DOTA datasets,and the m AP of the MMF-YOLO algorithm reaches 42.23% and 86.52%,respectively,which is 9.13% and 10.51% higher than the baseline network,and 1.57% and 2.06% higher than the Bi-YOLOX algorithm,respectively.(3)This paper designs and implements a portable small object detection system based on deep learning,which is mainly used for local and real-time detection of small objects in complex natural scenes.Firstly,Py Qt5 is used to design the UI graphical interface.Secondly,in order to smoothly transfer the algorithm model to the Jetson Xavier NX development board,the algorithm model is optimized.At the same time,in order to achieve real-time acceleration within the network,Tensor RT acceleration method is used on the Deep Stream platform.Finally,the detection system is able to perform near real-time detection of small objects.