Research And System Design Of Lightweight Deep Learning Target Detection Algorithm

With the development of computer vision,the target detection technology based on deep learning has made breakthroughs in computational accuracy and has been widely used in real life scenarios.However,the existing target detection network has the disadvantages of high computational cost and large model storage,which is not conducive to its deployment in devices with strict computational time requirements and low memory resources.Therefore,deep learning target detection technology based on li ghtweight has become a research hotspot for researchers at home and abroad.In order to make the network model smaller and faster without reducing the target detection accuracy,the model compression acceleration algorithm and deep learning target detection algorithm are studied in the thesis.The main contents of this thesis are as follows:(1)In this thesis,an improved lightweight Faster-RCNN target detection network algorithm is proposed.The algorithm reduces the network parameters and calculations by modifying the Faster-RCNN target detection network into a full convolution structure and removing the redundancy in the convolution kernel using a low rank decomposition algorithm.The algorithm is tested on the public Pascal VOC 2007 target detection data set.The experimental results show that the algorithm can reduce the model from 548.3MB to 38.2MB without reducing the accuracy,and compresses 14.35 times.The speed is improved 1.33 times.(2)Although the improved lightweight Faster-RCNN target detection algorithm has achieved good results,its speed and accuracy still need to be improvedTherefore,the fusion of SSD algorithm and MobileNet network are further studied to achieve more efficient targets.The experimental results on the Pascal VOC 2007 dataset show hat the model size can be reduced from 86MB to 14MB with a guaranteed accuracy of approximately 30 frames/second.(3)A two-class safety hat data set is made in this thesis,and a real-time light-weight deep learning target detection and recognition system is designed with the alg-orithm in(2).The system includes login system,algorithm selection,picture detection and recognition,video detection and recognition,real-time camera detection,log reco-rding,and setting thresholds.The average detection accuracy of the system on the helmet data set is 89.4%.Under the CPU of Intel Core i7-4770,the operation speed can reach 12 frames/second,which has the characteristics of real-time and low storage consumption.The algorithm can be transplanted to the mobile phone terminal.