Research And Application Of Real-time Object Detection Based On Lightweight Convolutional Neural Network

Object detection has always been a research hotspot in the field of computer vision due to its very important research and application value.Affected by factors such as lighting and environmental changes,traditional object detection based on artificial features has reached the bottleneck.And deep learning extracts features in images through autonomous learning,which has a very high adaptability to lighting and environmental changes.Therefore,the rapid development of deep learning has led to a major breakthrough in object detection.The object detection algorithm based on deep learning has excellent detection mAP,but usually the network models are very large,and the detection speed is low.It is difficult to apply to scenarios that require high detection speed or to deploy to mobile devices with limited computing power.In response to this problem,this thesis conducts lightweight convolutional neural network object detection research to ensure that the detection mAP remains unchanged while improving the detection speed.The main work includes:(1)According to the characteristics of the YOLOv3 network,designing model compression methods such as layer pruning,kernel pruning,and mixed pruning to reduce the size of the model and improve the speed of detection.The scaling factor learned from the BN layer is used as a measure of pruning,and a regularization term on the scaling factor is added to the original BN layer loss function.Performing sparse training,which is helpful for finding pruning targets of lower importance.For the characteristics of the network containing residual structure,targeted pruning strategies are adopted to ensure the integrity of the residual structure.The designed pruning methods are applied to the YOLOv3 model trained on the CCPD dataset.The detection speed is increased from 48 FPS to 303 FPS and the model size is compressed from 492.6 MB to 2.6 MB under the condition that the detection mAP is 98.7% unchanged.(2)Applying the lightweight YOLOv3 object detection method to the RoboMaster robot armor detection.In order to further improve the speed of detection,reducing the size of the network input image and the number of detection layers.Combining the method of shared memory to create a camera image acquisition process and an armor detection process and making the two processes run in parallel.The running speed of the final system on the CPU can reach more than 95 FPS,and the detection mAP reaches 95.8%.Aiming at the problems of large model and slow detection speed of deep learning object detection,this thesis introduces the method of network pruning on the basis of excellent performance of YOLOv3.Streamlining the network structure and greatly improving the detection speed while keeping the detection mAP unchanged.The validated method is applied to the RoboMaster robot armor detection,combining with other methods to further improve the detection speed and achieving real-time detection on CPU.This thesis contains 40 figures,17 tables and 53 references.