Research On Small Target Detection And Lightweight Method Based On Convolutional Neural Network

Object detection is one of the important research directions in the field of computer vision,its main task is to classify and locate the object in the image,and it is widely used in unmanned driving,industrial defect detection,intelligent monitoring and so on.Although object detection has made breakthrough progress,there are still many problems.On the one hand,the small object itself occupies a small proportion in the image and has fewer pixels.After multiple convolution and pooling of neural network,the small object information will be seriously lost.This will cause the target detection model can not obtain accurate and sufficient small object information,so that the detection effect can not reach the expected;On the other hand,in order to achieve excellent detection performance,some current advanced object detection models usually have a large number of functional modules,resulting in a very complex network model structure,which cannot be deployed on mobile devices with weak computing performance in practical application.As a result,these advanced object detection models can only stay at the laboratory level and cannot be applied in large-scale commercial applications.The main research contents and innovations of this paper are as follows:(1)A Recursive Attention-enhanced Bidirectional Feature Pyramid Network for Small Object is proposed Detection is proposed.Firstly,coordinate attention is embedded into the bidirectional feature pyramid network to build an attention-enhanced bidirectional feature pyramid network.The network integrates features at different scales to make the output features contain rich semantic and detailed information.Coordinate attention enables the network to focus on the object-related channels and positions in the feature map,improving the detection accuracy of small object detection.Secondly,a recursive structure is designed to feed back the output features of the attention-enhanced bidirectional feature pyramid network to the backbone network.The backbone network can find the hidden information in the image according to the results learned for the first time,and further improve the ability of the attention-enhanced bidirectional feature pyramid network to represent small targets.Experimental results show that the detection accuracy of the proposed method on PASCAL VOC,NWPU VHR-10 and RSOD data sets is improved by 2.65%,7.92% and 5.63%,respectively,compared with the original SSD algorithm.(2)A Lightweight Detection Network Based on Fusing Dual Attention and Depthwise Separable Convolution Ghost Bottleneck isproposed.First,we combine the DW-Ghost module with a fast connection that integrates double attention,and construct a fusing dual attention and depthwise separable convolution Ghost Bottleneck to lighten the YOLOv5 s backbone network,which reduces the number of network parameters and guarantees the ability of feature extraction.The feature extraction capability is ensured while the number of network parameters is reduced.Among them,the DW-Ghost module uses less computational deep separable convolution to generate cost eigenmaps,and based on these eigenmaps using low-cost linear operations to generate redundant ghost eigenmaps.Secondly,a cross-scale feature fusion network is designed to realize feature fusion at different scales,which not only simplifies the aggregation process of features at different layers,but also enables the fused features to obtain rich semantic and detailed information through lightweight channel attention and spatial attention.Experiments on PASCAL VOC and RSOD data sets show that this method has good performance.