| Object detection based on deep learning is an important branch in the field of computer vision,which is the basis of tasks such as instance segmentation,scene understanding and cognition.Its application scenarios involve auto-driving,face detection,intelligent security and other areas.In recent years,thanks to the improvement of computing power of hardware devices such as Graphic Processing Unit(GPU),many object detection networks with superior performance and complex structure have been proposed one after another.Although the detection accuracy has been continuously improved,there are weakness of slow speed and large network.It is difficult to deploy and apply on resource constrained devices,and not suitable for tasks with high real-time requirements.To solve the problems,based on the lightweight convolution network,the thesis proposes a novel object detection algorithm,called Lightweight Attention-based FCOS(LAFCOS),and the work realizes the balance between detection accuracy and speed.To overcome the difficult problem of small object detection,the thesis proposed a lightweight multiple receptive field module(LMRFB),which can the detection performance of the small and other objects.The main research contents of this thesis are as follows:(1)A object detection algorithm called lightweight Attention-based FCOS is proposed.Firstly,from the aspects of parameter,calculation,accuracy and speed delay,the thesis compares and analyzes the lightweight network obtained by the manual design and NAS.Based on FCOS algorithm,Mobile Net V2 and Efficient Net-Lite2 are chose as the backbone network of high-efficiency and high-performance versions of object detection model.Secondly,the thesis constructes Attention-based feature pyramid(AFPN)module,which the attention mechanism is added to the feature information transmission.The AFPN module realizes feature selection properly and highlight the information required by the task,which makes up for the performance degradation caused by the lightweight backbone network to a certain extent.Last but not least,adaptive training sample selection(ATSS)and generalized focus loss(GFL)are used during the training of LAFCOS model,which further improve the performance.Experimental results on the benchmark datasets show that our algorithm achieves higher object detection accuracy and speed than other lightweight network models,and even comparable to the heavyweight object detection algorithms.(2)A lightweight multiple receptive field block(LMRFB)is proposed.From the perspective of receptive field,hole convolution and serial cascade residual structure is introduced into LMRFB design,which obtains multiple receptive fields and increases the scene information around small targets.The thesis also adopts deep separable convolution to design LMRFB,which reduces parameters and makes it can be add to the network model without burden.Finally,this block is embed behind shadow feature map of the model.Experimental results on two benchmark datasets show that our lightweight multiple receptive field block can improve the detection accuracy of small targets and the whole. |
