Research And Implementation Of Small-scale Object Detection Based On Deep Learning

Object detection is a crucial research subfield in the field of computer vision,and has broad prospects in industries such as autonomous driving,video surveillance,and medical imaging.However,due to the small-scale objects occupies a small number of pixels in the image and lack effective features,small-scale object detection still has certain difficulties and challenges.This paper will conduct research and improve the small-scale object detection based on deep learning object detection algorithms.This paper is based on Faster R-CNN algorithm for research and improvement,a multi-layer feature fusion algorithm based on clustering and a Soft-NMS algorithm based on bilinear interpolation are proposed,and experiments are carried out.The multi-layer feature fusion algorithm based on clustering uses the multi-layer feature fusion method to make the extracted features have the characteristics of high resolution and high semantics;while The K-means clustering algorithm is used to generate the network layer module in the candidate region without prior knowledge to obtain a good anchor frame.The Soft-NMS algorithm based on bilinear interpolation uses bilinear interpolation on the region of interest layer module to make the position of the output suggested feature map more accurate;while in the classification layer module,soft NMS can better filter the detection frame and prevent false filtering of the detection frame.Based on the comparison experiments of the two improved algorithms proposed above,it can be seen from the test results of the experiments that the mean average precision of the two improved algorithms has a certain improvement compared with the Faster R-CNN algorithm.In conclusion,the research and implementation of small-scale object detection based on deep learning shows the feasibility and effectiveness of multi-layer feature fusion algorithm based on clustering and soft-nms algorithm based on bilinear interpolation in small-scale object detection.