Research On Lightweight Target Recognition Algorithm Based On Convolutional Neural Network

Target recognition is a crucial technology in computer vision.With the rise of deep learning,target recognition algorithms based on convolutional neural networks make up for the shortcomings of traditional algorithms in recognition accuracy and speed.However,the target recognition algorithms based on convolutional neural network have complex network calculations and large generation models,which occupy a large amount of computing resources and storage space of the device.When applied on embedded devices without GPU,the recognition accuracy and speed cannot reach the ideal effect.Based on the above problems,the thesis is based on the mainstream network YOLOv3,and studies from two aspects: improving the accuracy of algorithm recognition and lightening the recognition algorithm model.Aiming at the accuracy of target recognition,the thesis designs an improved highprecision target recognition algorithm CFPN?YOLO based on YOLOv3.CFPN?YOLO improves the original multi-scale fusion prediction method,adjusts the input size and adds a prediction layer to the fourth to last layer to perform more layer feature fusion,so that the network can obtain richer bottom-level details of the image.The algorithm introduces a consistent supervision loss function to reduce the loss of information between different prediction layers before the feature summation,and improve the recognition rate of the algorithm.Test the improved algorithm on NWPU VHR-10 and Aug-Fruits360 data sets,and compare them with other algorithms.Experimental results show that the recognition accuracy of the improved algorithm is 12.9% and 15.1% higher than that of YOLOv3,respectively.Aiming at the lightweight of target recognition algorithm,the thesis first designs the lightweight recognition algorithm LDS?YOLO based on CFPN?YOLO.LDS?YOLO combines deep separable convolution to improve the network convolution block,replacing the traditional convolution with 3×3 and 1×1 point convolutions,deepening the network depth and reducing model parameters.The algorithm introduces densely connected modules to replace residual modules,increases the correlation between different layers,and avoids the problems of overfitting and disappearance of gradients.The improved algorithm selects the Aug-Fruits360 dataset to complete the recognition model training,and tests it on the Raspberry Pi 3B+ platform without GPU to analyze its improvement and optimization degree in terms of target recognition accuracy,speed and parameter quantity.Experimental results show that the recognition accuracy of the improved algorithm is increased by 10.7%compared with YOLOv3.The average model parameter is only 8M,and the calculation amount is 125 MFLOPs.Then,by introducing a pruning operation to improve the weight of the feature channel,the model is further lightened.On the basis of ensuring the recognition rate,the model parameter amount is reduced to 3.6M,the calculation amount is 77 MFLOPs,and the test recognition rate loses less than 2.3% compared with the original model,the speed reaches 17 frames per second.Finally,by training and testing on other public data sets and comparing with mainstream recognition algorithms,the anti-overfitting ability and generalization ability of the lightweight algorithm model are analyzed,the effectiveness and feasibility of the lightweight target recognition algorithm based on convolutional neural network is proved.