Lightweight General Purpose Object Detection Algorithm

Object detection is an important research subject in computer vision field,which aims to distinguish classes and accurately locate various objects in images or videos by using algorithm models.Object detection technology is widely used in automatic driving,intelligent Internet of Things,national defense and military,monitoring and security and other fields.At the same time,object detection is also the upstream task of many other computer vision tasks,such as pedestrian detection,video description and generative AI,etc.The performance of object detection technology also affects the performance of downstream tasks,which has a wide range of research and application in academia and industry.With the explosive development of deep neural networks in recent years,great breakthroughs have been made in the field of object detection with the help of deep learning technology,and a paradigm with convolutional neural networks as the main algorithm framework has been established.At present,the general object detection model can perform various object detection tasks well in the laboratory scene,but it is not lightweight enough.The laboratory model is difficult to carry on the edge equipment in the actual detection scene.In addition,various targets as object detection objects have differences in physical characteristics,appearance,light and shadow,posture,scale and other aspects,as well as in the camera imaging process there are objective conditions such as vibration,illumination,which also enhance the difficulty of object detection task.In view of the problems described above,this paper,based on the deep convolutional network model of object detection,improves the existing neural network and designs a new neural network to achieve the lightweight of object detection network specifically and effectively.The specific research of this paper can be introduced from the following three aspects:(1)The mainstream algorithm has its own lightweight version,but its accuracy is seriously degraded.Based on the YOLOv4-Tiny algorithm,this paper proposes an improved method based on attention and feature fusion.First,we modified the activation function in the backbone network and introduced the ECA attention mechanism,and modified the convolutional layer of the backbone network into the ECA-resblock module.The attention mechanism made the model focus on key targets and give higher weight to interested targets.The traditional SPP module is improved,and the void convolution is used to replace the original maximum pooling layer structure,and the bottleneck block structure of Res Net is used to bring multi-scale feature information to the main trunk network.CSATT-PAN structure is used to replace the original neck network,and more context information is provided to the model through the feature fusion of high-level semantics and low-level semantics,which improves the accuracy of the model.(2)According to the network design idea of light backbone network and heavy neck network,a lightweight universal object detection algorithm is designed in this paper.In this chapter,YOLOv4-Tiny is taken as the baseline network.Firstly,Ghost module is used to replace the original backbone network of the baseline model,which greatly reduces the number of parameters and calculation.The channel mixing mechanism in Shuffle Net is added to the input side of the backbone network,so that the input information contains cross-channel information.The DC-SPP module is added to the bottom output side of the backbone network to integrate the richest spatial information at the bottom.Position loss based on CIo U makes the prediction frame more compact and fit to the real target.Finally,inspired by Giraffedet,a neck feature fusion network is designed,which adds cross-layer connections between upper and lower layers and residuals hop connections between front and back layers compared with previous networks.The high density exchange of spatial information is realized.Finally,the experimental results also reflect the effectiveness of our improved model.(3)Aiming at the problems existing in anchor-based algorithm,the domain algorithm FCOS based on anchor-free algorithm is tried.Based on the baseline algorithm FCOS object detection algorithm,this chapter attempts to design the object detection algorithm under the anchor-free paradigm to further lighten the FCOS algorithm.Firstly,the backbone network of the algorithm is replaced with the lightweight network Mobile Netv2.Due to its deep separable convolution and inverted residual structure bottleneck block,the number of parameters and calculation amount can be greatly reduced.Then,the deep convolution filter is used,and the linear bottleneck module is used to project the features back to low latitude.It preserves most of the manifold of interest in the data space.In the neck network,P6 and P7 layers,which contain less information,are discarded,further reducing the number of parameters and calculation.Then,aiming at the object occlusion problem in the image,a spatial context module is proposed,aiming at using the global spatial context information to assist the model to make better prediction in the absence of complete image information,and combining with the integration of local information to improve the performance of the model.In summary,this paper explores the lightweight of target detection algorithm from three aspects,and carries out comparative verification on different data sets,which proves the effectiveness of the proposed method.