Improved Algorithm Of Object Detection Based On One Stage Network Model

Object detection is a significant topic in the field of digital image signal processing,and it is also the premise of other advanced visual tasks.It is widely used in automatic driving,remote sensing detection and other fields.Traditional object detection algorithms have low accuracy and poor real-time performance.Object detection algorithms based on deep learning have improved this phenomenon to a certain extent,but there are still problems such as difficulty in detecting small objects.In order to reduce the adverse effects of such problems on the algorithm and improve the detection performance of the algorithm,this paper conducts an in-depth study on the object detection algorithm based on the one-stage network model-SSD algorithm,and proposes corresponding improvement methods for the shortcomings of the algorithm.The specific content is as follows:(1)Aiming at the limitation of SSD object detection algorithm based on one-stage network model and the difficulty of small object detection in input image,an object detection algorithm combining feature pyramid network and attention mechanism is proposed Firstly,the idea of feature pyramid network is used to fuse the multi-scale feature map of SSD algorithm to lay the foundation for subsequent detection;secondly,attention is added to the last two layers of feature map after fusion to improve the weight of region of interest;finally,after comparing the two optimization algorithms,the network is trained to complete the object detection task.(2)In the SSD object detection algorithm based on the one-stage network model,because the edge object features are not obvious,the basic network downsampling operation causes the image resolution to be reduced,and the local information is missing.At the same time,due to the insufficient receptive field of the algorithm,it is difficult to detect the object at the edge of the image.Propose a multi-scale feature fusion object detection algorithm based on hole convolution.Firstly,the conv4?3convolutional layer and the previous two standard convolutions are replaced by the hole convolution with different expansion rates to better mine the global information in the image;secondly,the feature maps of different scales are combined with deconvolution to perform skip feature fusion to complete the high-level The feature map is mapped on the underlying feature map;finally,attention is added to all the feature maps,and the information on the feature maps is integrated to further improve the detection performance.(3)Model training and performance comparison between the improved algorithm and several other mainstream algorithms on public data sets.Experimental results show that compared with other algorithms,the improved algorithm improves the original algorithm's difficulty in detecting small objects and edge objects,still retains a faster detection speed.