Research On Small Target Detection Algorithm Based On SSD

With the rapid development of deep learning and the gradual improvement of computer equipment,the current target detection techniques are mostly based on the deep learning approach,while the traditional target detection technology has been gradually lonely,in more and more fields appeared the figure of the target detection technology based on deep learning,such as:autonomous driving,face recognition,image retrieval,industrial parts inspection,etc.However,in real life,there are often problems when using deep learning-based target detection technology,such as: mutual occlusion between targets,target objects are too small to be recognized,network model parameters are too large to meet real-time requirements,etc.There is still a lot of room for optimization and improvement of the target detection algorithm,which is a key research issue in the academic circle,and has great practical significance and value.Compared with SSD algorithm,the detection accuracy of EFSSD for small targets has been significantly improved,while meeting real-time requirements.The main research contents of this paper are as follows:This paper takes SSD in single-stage target detection algorithm as the main research object,analyzes the advantages and disadvantages of SSD algorithm,optimizes SSD algorithm,and proposes EFSSD algorithm on this basis.(1)Research on SSD target detection algorithm.To solve the problem of poor detection of occluded targets,the cross-entropy loss function was replaced with Focal Loss,and soft-NMS was introduced to improve the NMS of the original SSD algorithm,and its feasibility was verified on the Pascal VOC2007 test set.The experimental results show that the detection accuracy of the improved SSD target detection algorithm in this paper has been improved.(2)Aiming at the problem that the SSD algorithm is not effective for small target detection,this paper proposes the EFSSD target detection algorithm.First,the ECANet attention mechanism is added on the basis of the optimization of the SSD in Chapter 3.Through one-dimensional convolution cross-channel interaction,dimensionality reduction is avoided,and cross-channel interaction information is effectively captured.The ECANet module is a lightweight network with small parameters but significant improvement effects.Secondly,in this paper,conv4＿3 is concatenated with conv6＿2 after upsampling,and nconv4＿3 convolutional layer is obtained,which is sent to detection together with conv7,conv8＿2,conv9＿2,conv10＿2,and conv11＿2processed by ECANet.Finally,a comparative experiment shows that the EFSSD proposed in this paper has higher accuracy and a better effect on the detection of small targets.