Wearable Object Fast Detection Method Based on Fine-Grained Feature Extraction
In order to solve the problems of slow recognition speed,poor anti-interference ability,false detection and missed detection caused by size change,light brightness and darkness,partial occlusion in visual detection of human wearable targets,especially the discrimination of similar targets,a rapid detection method of wearable targets based on fine-grained feature purification was proposed,which added jump connections in the CBAM structure to obtain feature maps with dual characteristics of space and channel,while retaining richer original information.The self-attention mechanism and convolutional neural network were integrated to improve the perception of global information in the backbone network,a feature pyramid network was designed to facilitate multi-size object detection.The shallow position information and deep semantic information were extracted at the same time,which greatly improved the detection accuracy.Ablation experiments were carried out on the MS COCO dataset to verify the influence of each module on the network,and the effectiveness and advancement of the proposed method were proved in comparative experiments.On the MS COCO 2017 dataset,the AP50 value reached 60.5,the AP value reached 35.0,and the detection speed was 5.7 ms.Compared with YOLOv5s,the detection speed was increased by 18.6％,the computing power requirement was reduced by 33.3％,and the number of parameters was reduced by 16.7％while the accuracy was similar.The AP value of this method on the high-altitude seat belt dataset reached 62.5,which was better than the mainstream deep learning object detection methods.
deep learningmachine visionattention mechanismfine-grained object detectionwearable object detection
万方数据
维普
