Optimization Method Of A Lightweight Deep Learning Model And Its Application In Object Detection

In recent years,significant advances have been made in fields such as natural language processing,speech recognition,and computer vision.Benefit from this,great progress in artificial intelligence has been seen.Artificial intelligence is widely applied in various professions,such as self-driving cars,robotics,unmanned aerial vehicles,etc.Their navigation and obstacle avoidance systems are highly depended on real-time object detection algorithm.With the emergence of large-scale image datasets and highperformance GPU,unprecedented achievements have been made in image classification and object detection.Various models and algorithms based on convolutional neural networks(CNN)emerged,making CNN the dominant solution to object detection.However,deep learning models are usually impractical to be deployed,poor real-time performance and high energy consumption are their common problems.In this paper,an object detection system that can be deployed on embedded platforms has been built with deep learning methods,and a lightweight convolutional network architecture has been proposed to optimize its structure and accelerate the inference process.The main contents of this paper can be summarized as follows:Firstly,a lightweight convolutional neural network architecture is designed to extract features for detection.Inspired by Resnet,Shuffle Net,Darknet and other architectures,the feature extraction network is built using depth-wise separable convolution and residual structure.Combined with the one-stage YOLO detection algorithm,a model family called Shuffle Detector is constructed to meet the application requirements of different scenarios.Either model with higher accuracy or higher efficiency can be chosen according to the practical demand.Secondly,two dataset of nighttime road scenarios are built to provide support for subsequent research work.Image data were captured using FLIR Zenmuse XT thermal camera,then manually annotated and reviewed.According to the specification of PASCAL VOC2007,two datasets for object detection were constructed,corresponding to red-hot and white-hot modes respectively.These two datasets are used to test the performance of Shuffle Detector.Thirdly,attention mechanism and inference optimization are used to optimize the trained model and then deployed on Jetson TX2.In this paper,channel attention mechanism and spatial attention mechanism both are applied to optimize the network structure and enhance its ability of feature extraction.During deployment,the computational graph is simplified to reduce the calculation and scheduling overhead.At the same time,half-precision inference is used to run the model with 16-bit floatingpoint number,which further improves its efficiency.Through experiments,Shuffle Detector can provide better performance on both two datasets with less params and MACs.The model with the least parameters--Shuffle Detector Half can achieve 128 FPS on Jetson TX2 when testing.With higher real-time performance,it is more suitable for tasks that requiring fast object detection.