Research Of Lightweight Small Object Detection Algorithm Based On Deep Learning

With the development of deep learning technology,the research and application of deep learning have successfully solved many academic and engineering problems.In the object detection task,from two-stage to one-stage,and then to anchor-free detectors,the average precision is constantly improved.However,there are still two drawbacks.Firstly,the average precision of small object is low.Secondly,with the increase of the average precision,the number of operations and parameters are also increasing,result in a larger model size and a slower detection speed.These problems limit the application of object detection algorithm in mobile robots,self-driving vehicles,smart phones and other mobile platforms or embedded devices.To solve these problems and promote the application of object detectors.In this paper,a new algorithm named LMDet-S(Lightweight Multi-scale Detector for Small Object)was proposed.The algorithm is mainly divided into two parts,one is the Lightweight feature fusion network Lite-PAN,the other is the Lightweight detection Head LAF-Head.To verify the performance of the algorithm,many experiments are carried out.Firstly,the experiments on the standard dataset COCO show that,compared with the SOTA algorithm YOLOv4-Tiny,the FLOPs of the proposed algorithm are reduced by 42% and the params is reduced by 34% at the same resolution.Meanwhile,the mAP of the proposed algorithm is improved by 8.6%,and the APs is improved by 2%.Then we applied our algorithm on Tiny Benchmark such as Tiny Person.Based on the “image cutting” and “scale match” strategies,we use single scale training and single model,our algorithm achieves better performance.Finally,the deployment experiment is conducted on a vivo-iQOO-Z1x mobile phone.The real machine test shows that the speed of our algorithm based on ShuffleNetV2 backbone can achieve 30 FPS and satisfy the real-time standard.