Research on Intelligent Video Detection of Small Targets Based on Deep Learning Intelligent Algorithm

Table 2

Target detection methods combined with lightweighting strategies.

Method name	Year	Basic introduction	Performance
CSPNet	2019	Starting from the perspective of network architecture, we use cross-stage feature fusion to optimize the repetitive gradient information in the network to achieve lightweighting	In the same environment, the computation is reduced by nearly 30% and the accuracy is improved by 2% compared to yolov3
YOLO Nano	2019	Design PEP macro-architecture by human-machine collaborative design strategy, combined with fully connected attention module for embedded environment to significantly reduce the amount of computation, but only in embedded environment	Achieve an average accuracy of 69.1% on the VOC2007 data set and a model size of only 4 MB
ThunderNet	2019	Based on ShuffieNet v2, compressing RPN module, proposing context enhancement module, using 1 × 1 convolutional compression channel to achieve feature fusion effect while reducing computational cost, and introducing spatial attention mechanism to optimize feature distribution to reduce computational effort	Obtain 19.1% AP on the COCO data set, which is similar to the accuracy of SSD using MobileNet, but nearly 5 times faster and significantly cheaper to compute
RefineDeLite	2020	A new backbone network Res2NetLite is proposed for the detection task, ensuring the same number of input and output channels, focusing on optimizing the loss function and training strategy	The AP value of 26.8% is achieved on the COCO data set, and it can reach 29.6% with its proposed training strategy, which is the best lightweight network at present