Lightweight Object Detection Algorithm Based On Improved SSD

Object detection is a computer vision technique used to locate object instances in an image or video.Object detection algorithms are usually combined with convolutional neural networks in deep learning to produce meaningful results,and their outstanding performance has become the mainstream of artificial intelligence.With the development of object detection algorithm based on convolutional neural network,it consumes more and more computing resources and occupies more and more storage space,which also raises the requirement of equipment.Existing mobile devices,embedded devices and other computing storage space is insufficient,the real-time algorithm is poor,so it is very necessary to design an efficient,stable and small storage lightweight target detection algorithm.Firstly,this paper proposes Mobilenet V3-SSDbif PN model,which uses a lightweight network combining Mobile Net V3 and SSD,and then adds a bidirectional feature pyramid model to fuse features,which effectively improves the network detection accuracy.The model was trained and validated using Pascal VOC2007 and2012 datasets,and compared with SSD,Mobile Net series,Tiny-YOLOV3 and literature [56] algorithms.The experimental results show that the model reduces the memory space,and the detection accuracy m AP and the detection speed FPS reach73.65%and 28frames/s respectively,which proves that the model has good detection ability.After that,this paper improves the Mobile Net V3-SSDBifpn algorithm and proposes an object detection model based on SCE algorithm,which is also trained and verified by PASCAL VOC2007 and 2012 datasets.The algorithm is mainly improved from three aspects: bounding box localization regression loss,non-maximum suppression and channel attention mechanism.The specific work is as follows:(1)In order to improve the accuracy of target detection and positioning,an improved Complete Intersection over Union(CIo U)loss is used as the positioning loss in the network loss,which adds aspect ratio and center point distance.And this is used as the penalty term of the bounding box coordinate predicted loss function to improve the regression accuracy of the algorithm loss function.It not only accelerates the model training speed,but also improves the detection and positioning accuracy of the model.(2)In order to solve the problem of missing detection caused by directly deleting too many overlapping detection boxes in traditional Non-Maximum Suppression(NMS),this paper proposes Soft-NMS(Soft Non-Maximum Suppression).The NMS algorithm is modified without adding additional parameters,which has the same efficiency as NMS,and it does not need additional training,so it has the characteristics of plug-and-play in object detection based on convolutional neural network.(3)Aiming at the problem that the dimension reduction channel attention caused by SENet leads to the adverse mutual learning of feature channels,the ECA-Net network that does not change the dimension for cross-channel information interaction is used to replace SENet.The global average pooled GAP is used to obtain the global receptive field,in order to obtain deeper semantic features,reduce the complexity of the model,provide performance,and effectively improve the accuracy of the model.To sum up the above three aspects,the experimental results show that after integrating the CIo U,Soft-NMS and ECA-Net methods,the memory size and FPS of the SCE object detection algorithm reach 75.37%,29 MB and 33.1frames/s,which is5 MB lower than that of the Mobile Net V3-SSDBifpn algorithm,and 1.72% and 5.1higher respectively for m AP and FPS,making it more suitable for mobile and embedded devices.