Low Resolution Image Object Detection Based On Convolutional Neural Network

The high-resolution image has a clear picture quality with abundant information,but in practice,most of them are low-resolution images with little information and small size.Low-resolution images mainly lack high-frequency details,so the object detection of low-resolution images has always been a difficult problem.Low resolution image reconstruction technology can reconstruct the high-frequency information of low-resolution images,so it is possible to solve the problem of low-resolution images object detection.This paper focuses on the object detection algorithm of low resolution image based on deep learning.The main work is as follows:This paper proposes the Multi-scale Densely Connection Network(MSDN)based on the Multi-scale Residual Network(MSRN),which is optimized as follows: In order to make full use of the structural similarity between high-and low-resolution images,and obtain better high-frequency information,we use dilated convolution instead of ordinary convolution to increase the receptive field of the convolution kernel and obtain higher-level semantic features.In order to fully utilize the features of each stage,we use densely connection structure to fuse low-level features and high-level features.The fusion has alleviated the problem of low-level features easily disappearing in the deep network,and densely connection structure has alleviated the problem of gradient vanished.In order to solve the problem of low-resolution image object detection,a network combining low-resolution image reconstruction model and object detection model(MS-YOLO)is proposed in this paper.In the proposed network,the input low-resolution image is reconstructed first,and then the reconstruction result is used as the input of the object detection algorithm.In order to enrich the available features of object detection,a multi-scale feature mapping module(MSFM)is proposed.The Conv-Leaky Re LU-BN combined module performs feature mapping,and fuses the mapped multi-scale features with the features extracted by the object detection algorithm to increase the accuracy of object detection model.In this paper,MSDN and other classic reconstruction algorithms are compared on objective and subjective indicators respectively,focusing on the high-frequency details of the reconstructed image,and analyzing the model complexity of MSDN and MSRN.The experiments results show that the MSDN network is concise And the characteristics of better reconstruction effect,at the same time,MS-YOLO and other object detection algorithms are compared in low-resolution images,focusing on the improvement of object detection accuracy.The experimental results show that MS-YOLO can effectively solve low-resolution image object detection problems.