Research On Fast Target Detection Method Based On Lightweight Network And Feature Fusion

As a key research content in the field of computer vision,object detection has been widely used in various life scenarios such as traffic travel,security monitoring,and health care.At present,traditional object detection methods have low detection precision,and some deep learning based object detection methods significantly improve the accuracy of detection.However,due to the deeper and more complex convolution neural networks,the detection speed is difficult to meet the needs of practical applications.With the emergence of deep learning model in embedded and mobile deployment,how to improve the detection speed and reduce the size of the model while ensuring the accuracy has become the research focus in the field of object detection.In this paper,we proposed a fast object detection method based on lightweight convolution neural network.By designing an efficient and simplified convolutional neural network,the speed of detection can be significantly improved.Moreover,aiming at the difficulties of detecting small objects,occluded objects and dense objects,we proposed MFFD: Multi-scaled and Feature Fusion Object Detector,using a lightweight convolution neural network.The major innovative works are as follows:(1)A fast object deteciton method based on the lightweight convolution neural network was proposed.we introduce a class of efficient network architecture named ThinNet mainly for object detection applications on memory and computation limited platforms.The new architecture is based on two proposed modules: Front module and Tinier module.The Front module reduce the information loss from raw input images.The Tinier module use pointwise convolution layers before conventional convolution layer to decrease model size and computation.A series of comparative experiments shows that ThinNet can ensure the accuracy while requiring smaller storage space to run at a faster speed.(2)Multi-scaled and Feature Fusion Object Detector was proposed.By fusing feature maps of different level and detecting in multi scaled feature maps,the detection performance of small objects,occluded objects and dense objects have been greatly improved.We design two ways to concatenate features from different layers with different scales,aiming to add contextual information.Experiments on challenging datasets show that MFFD outperforms many popular methods.In the exploratory experiments of embedded devices and low-power devices,MFFD can achieve more than 100 fps,which is fully applicable to real-time object detection applications in real scenes.(3)This paper designed and implemented a fast object detection system based on deep learning.The system was divided into three functional modules: data selection module,detection process module and result save module.The system can display the result images and varous evaluation indicators.Moreover,This system supports saving result information and batch automatic detection.