Research On MobileNet-SSD Target Detection Algorithm Fusion Attention Mechanism

Object detection algorithm is the core foundation of computer vision,and has important research and application significance.With the development of deep learning,the field of computer vision also plays an important role in various fields.However,because traditional target detection algorithms usually use hand-designed features for classification,they are not compatible with the diversity of targets,and most of the deep learning target detection algorithms use a large amount of model parameters,slow calculation speed,and cannot integrate well.in mobile or embedded devices.Therefore,designing a target detection network with higher accuracy and faster speed has become an urgent problem to be solved in both academic and practical applications.In response to this problem,this paper proposes a Mobile Net-SSD model fused with attention mechanism,which adopts the lightweight network model Mobile Net with depth separable convolution,and combines the SSD detection network for object detection.At the same time,the attention mechanism module is integrated into the model to improve the accuracy of the detection model.The comparative experimental results show that the improved Mobile Net-SSD network reduces the number of parameters by 33 times compared with VGG,and after incorporating the CBAM attention mechanism module,the detection accuracy is increased by 1%.In order to further improve the accuracy,instead of the Smooth＿L1 algorithm,the improved IOU loss function is used as the position loss function,and the positional relationship between the predicted frame and the real frame is comprehensively considered to optimize the learning strategy and improve the training effect.And the pre-selection frame processing method is improved,and the Soft-NMS algorithm of frame fusion is used for the de-overlapping frame algorithm to optimize the detection performance of multi-target overlap.The experimental results show that the detection accuracy of the improved algorithm model is increased by 3%.For mobile devices or embedded devices,it does not require a large amount of computation to meet the needs.The experimental results on the Pascal VOC2007 and VOC2012 datasets show that the model achieves good performance in both recognition accuracy and detection speed.The detection speed can reach 60.9fps on GTX1080,and the m AP accuracy can reach 73.6%.For mobile and embedded devices,the model is compressed,the target detection application software is designed,and experiments are carried out.The model parameters in this paper are small,the calculation speed is fast,and the recognition accuracy is high,which can better realize the real-time detection of mobile terminals and embedded systems.