Research On Small Object Detection Algorithm Based On Deep Learning

With the rapid development of deep learning technology,object detection techniques are also constantly improving and have been widely used in various industries.In many application scenarios,small objects are an important type of target that requires accurate identification and detection.Therefore,researching small object detection is of great significance.However,small object detection has always been one of the challenges in the field of object detection.Due to the small size of the object itself,the lack of features,susceptibility to background interference,neural networks have difficulty effectively extracting features of small objects,resulting in low detection precision.To improve the detection precision of small objects in complex scenes and large-scale spans,we propose small object detection algorithms based on feature enhancement and multi-scale,respectively,as described below:(1)A small object detection algorithm based on feature enhancement is proposed to improve the detection precision of small objects in complex scenes.The network neck adopts a bidirectional sampling feature pyramid network,which introduces deeper and shallower feature layers and constructs two feature aggregation paths to fuse richer semantic and spatial information.At the same time,adjacent three layers of feature maps are adaptively fused to enhance the features of small objects on the feature map.The detection head adopts a decoupled head with attention,which reduces conflicts between information at different levels through a three-branch decoupling structure.Furthermore,the attention mechanism is improved and introduced into the three-branch decoupled head to enhance the features of small objects and suppress background information.(2)A multi-scale small object detection algorithm is proposed to improve the detection precision of small objects with large-scale spans.The network body adopts a scale decoupled networks,which uses dilated convolutions to provide different receptive fields,and detects targets of different scales through a multi-branch detection architecture to reduce the interference in the feature extraction process of small objects.At the same time,a channel is constructed to aggregate from the deep network to the shallow network,enriching the semantic information on the small object detection layer.In the backbone network,an upsampling aggregation module is used to introduce feature fusion into the downsampling process,reducing information loss and enhancing the effect of feature fusion.Meanwhile,a multi-level recursive structure is used in combination with residual modules to enhance the network’s representation ability for small objects.In summary,we propose small object detection algorithms based on feature enhancement and multi-scale to address the problem of small object detection in complex scenes and large-scale spans.By introducing different network architectures and modules to enhance the network’s representation ability for small objects,the detection precision of small objects is improved.By conducting experiments on the remote sensing dataset RSOD and the conventional dataset VOC,the effectiveness of the two proposed algorithms for small object detection is verified.