Design Of Lightweight Anchor-free Object Detection Network

The development of computer vision technology promotes the development and progress of automatic driving,security monitoring and other fields,among which the object detection algorithm plays an important role.At present,the object detection based on deep learning is widely used.In the detection process,anchor mechanism is often used to improve performance.However,the anchor needs to introduce super parameter and has a large calculation amount and slow inference speed.However,the algorithm without anchor points avoids the disadvantages of anchor mechanism,but it needs large feature extraction network to ensure the accuracy,which leads to large model size,and cannot be deployed when facing a mobile terminal with limited computing power.Therefore,it is of great significance to build a lightweight anchor-free object detection algorithm.Based on the framework of Center Net network,a lightweight anchor-free object detection network Center Det SQ is proposed in this thesis,which maintains the speed advantage of the anchor-free detection network,and compresses the model size.The structure of detection network consists of three parts,which are feature extraction network,key point extraction module and output decoding.In the process of detecting network design,the improvement and optimization work are included.Firstly,the solution to the problem of key point mismatch in the anchor-free detection network is given;Secondly,based on the improvement of Center Net network structure,the method of feature extraction is optimized,and the detection performance is greatly improved;In addition,the lightweight module is adapted to feature extraction network,which greatly compresses the size of network model and accelerates the inference speed;Finally,the model training method is optimized,and the training accuracy is further improved by the improvement of data augmentation and activation function.In this thesis,a deep learning training environment through Pytorch is built and the detection model is trained on MS COCO dataset,which is a large object detection dataset proposed by Microsoft.The model size is 16.5% of the size of Center Net model,the inference speed is increased by 6.57 times,the detection accuracy is 35.4m AP,which exceeds the same level detection method of 4.2m AP,and the average test speed of each picture on one Tesla V100 GPU is 42 ms.The Center Det SQ can be deployed on the embedded platform Raspberry PI 4B used in this thesis.The objects can be accurately identified and detected,thus promoting the application of anchor-free algorithm in the embedded platform.