Research On Lightweight Convolutional Neural Networks For Target Recognition In The Auxiliary Waistcoat For The Blind

There is the largest group visual impairment of the world in China,but they are rarely seen in daily life except in special places such as schools and massage shops for the blind.Therefore,the society began to pay extensive attention to the blind people’s travel,life and other issues.Due to the continuous improvement of hardware technology,the data processing capability of embedded devices becomes stronger rapidly,which provides the technical basis for the research of wearable intelligent devices.It is of great social significance to research and develop wearable intelligent devices for the visual impairment.At present,the mainstream target recognition algorithms have the problems of complex network structure,large amount of computation and high degree of model fragmentation,which can not realize real-time data processing on embedded devices.After entering the 5G era,researchers upload the data collected by embedded devices to the cloud,process the data in the cloud computer,and then send the results back to the embedded devices,so as to realize the transplantation of algorithm in a disguised form.Although this method can ensure the recognition accuracy and real-time performance,it will bring inconvenience to users once the device suffers from network impact and signal loss.In particular,such problems in aid equipment will directly threaten the safety of the blind and cause irreparable consequences.Firstly,aiming at the above problems,this paper designs the hardware platform of the auxiliary waistcoat for the blind,studies the basic structure and working principle of convolutional neural networks,deduces relevant optimization formulas of convolutional neural networks,analyzes several common lightweight convolutional neural network models.Secondly,according to the principle of target recognition algorithm and the design principle of lightweight convolutional neural networks,lightweight convolution module,lightweight feature extraction module and feature enhancement module were designed.Then,based on Feature Pyramid Networks(FPN),the combination of algorithm adopts top-down FPN and bottom-up FPN to construct a multi-scale feature fusion network model.After that,the neural network is built and trained on the PC platform.Head of the original model was turned into a Decoupled Head for a comparative experiment.It was found that it could not only accelerate model training,but also improve the accuracy of model recognition.In addition,it was also found that Anchor Free was used to reduce the number of parameters and computation amount of the model.The improved Lightweight target recognition networks based on decoupled(LTDNet)model is applied to target recognition in the auxiliary waistcoat for the blind.Finally,experiments were carried out on different hardware devices to compare the performance of the proposed algorithm with the current mainstream target recognition algorithm.The performance of the proposed algorithm was measured by FLOPs,Params,mAP,frames per second(FPS)and the test data curve of three image sizes.The performance of LTD-Net model of embedded device transplantation was evaluated by four hardware performance indexes,such as running time and energy efficiency ratio.