Research On Single-stage Target Detection Algorithm Based On Deep Learning

As one of the basic topics in the field of image processing and computer vision,target detection tasks have a wide range of applications in image retrieval,video surveillance,face detection,and human-computer interaction.The traditional target detection method firstly performs feature extraction on a given picture and then performs classification selection.Therefore,the quality of feature extraction plays a key role in the performance of target detection.Compared with the traditional target detection method,the deep learning-based target detection method can adaptively learn a better feature extraction method through a large amount of data training,which can better handle complex scenes.At present,there are two types of target detection methods based on deep learning: the first one is a two-stage target detection method,and the other is a single-stage target detection method.The two-stage target detection method first needs to divide the image into regions and then perform detection.The single-stage target detection method is to directly output the detection result after inputting the picture.The two-stage method has certain advantages over the single-stage method,but the process of candidate region partitioning is computationally intensive and cannot meet the requirements of real-time.Therefore,this paper studies the single-stage target detection method.The work of this paper and the main research results mainly include the following aspects:(1)There is not much connection between different scale feature layers involved in prediction in SSD target detection algorithm.In order to improve the relationship between contexts,this paper integrates the information of different scale convolution feature layers based on the fusion attention mechanism.The fusion attention mechanism is a new attention mechanism.It firstly scales the feature layers of different scales by scale operation,and then introduces the context information into the SSD algorithm through convolution and matrix multiplication.On the PASCAL VOC dataset,experiments have shown that the improved SSD algorithm that introduces context information has a better effect on m AP performance than the original SSD algorithm.(2)The VGGNet16 network structure adopted by the feedforward network of the SSD algorithm.Although VGGNet16 is one of the more classic models,the shallower effect of the neural network layer needs to be improved.Therefore,this paper replaces the feedforward network of the SDD algorithm with a residual network with a better layer number and constructs a new model structure Res Net-SSD.It is proved by experiments that although VGGNet16 is replaced by the residual network,the effect will not improve,and its m AP performance will decrease.However,after adding the fusion attention and sharing prediction module,the residual network is used as the model of the feedforward network.The performance is better than the model accuracy performance with VGGNet16 as the feedforward network.(3)In the target detection task,this paper proposes a shared prediction module(Shared PM)to increase the receptive field range of different feature layers by using cavity convolution(expansion convolution),and at the same time The amount of parameters and the efficiency of the calculation,this paper uses the operation of shared convolution parameters.Experiments show that adding a shared prediction module based on the improved SSD algorithm can achieve better m AP performance on the PASCAL VOC dataset.