Research On Small Object Detection Algorithm Based On Deep Learning

With the rapid development of deep learning in object detection algorithms,small object detection plays an important role in fields such as autonomous driving and security monitoring.This paper mainly combines object detection algorithms and superresolution reconstruction algorithms to study small object detection and analyze the shortcomings of the algorithms.To address issues such as low resolution and missing semantic information of small objects,based on deep learning object detection algorithms and super-resolution reconstruction algorithms,we enhance the features of small objects and improve detection accuracy using the COCO dataset,PASCAL VOC dataset,and wildlife dataset as research objects.The research content of this paper is applied to the winter Bai Shan wildlife detection project.The main research content and innovations of this paper are as follows:(1)For the problem of missing semantic information of small objects,we propose a feature-enhanced and deep-level fused object detection algorithm(FEDet)based on feature optimization.This algorithm improves the SSD algorithm by using the spatial channel feature enhancement(SCFE)module and the deep-level feature pyramid network(DFPN).The SCFE module optimizes the feature layer based on the local spatial feature enhancement and global channel feature enhancement mechanisms,emphasizing the detailed information of the feature layer.DFPN improves FPN by using the residual spatial channel enhancement module,enabling deep-level feature fusion of different scale feature layers to improve object detection accuracy.At the same time,this algorithm adds a sample-weighted training strategy during the training stage,focusing on training samples with good localization and high confidence.The experimental results show that this algorithm ensures a speed of 79.7% detection accuracy on the PASCAL VOC dataset while detecting small targets with an accuracy6.5% higher than the original SSD on the COCO dataset.(2)For small objects with low resolution and blur,we propose a novel algorithm named MCRH-SRGAN,which is based on multi-level encoding and decoding and residual space-variant dilated convolution(RSAC)for enhancing the image understanding capability without increasing the number of parameters and computational complexity.By extracting detailed features using RSAC,we generate super-resolution images that are closer to the high-resolution images.Additionally,we use a multi-level encoding and decoding module(MCM)as the discriminator of the GAN to learn local texture information,which improves the discriminative ability and stabilizes training.Experimental results show that our algorithm achieves PSNR and SSIM of 27.26 d B and 0.8118,respectively,on the wild animal dataset.(3)In order to verify the small object detection effect of the proposed algorithm,this chapter combines FEDet and MCRH-SRGAN into a deep network algorithm called DNASO,and applies it to the wildlife dataset.DNASO consists of a main framework and a sub-framework.In the main framework,the image is input to FEDet,which detects all objects in the image and transfers small objects with poor recognition results to the sub-framework.In the sub-framework,MCRH-SRGAN is used to perform superresolution reconstruction on the object,generating a super-resolution image for secondary detection.The experimental results show that DNASO can effectively improve the detection accuracy of small objects.