Research On Lightweight Target Detection Algorithm For Similar Small Target Detection Tasks

The object detection task is to find out the objects of interest in the image by using the algorithm and determine their position and category,which is one of the core problems in the field of machine vision.Up to now,many sub-problems in the field of target detection have been fully studied,but there are still some problems to be solved,such as similar small target detection.Similar small target detection task refers to the recognition and differentiation of multiple targets with similar or even identical appearance.Because it is impossible to distinguish by the differences in appearance features such as the color and shape of the target,and the target occupies a small part in the image with few details,it is more challenging than the general target detection task.Similar small target detection has certain applications in industrial production and other scenarios.For example,when AR technology is used to assist the assembly process of bolt sets in important parts,multiple bolts of the same type in different positions need to be distinguished.Therefore,this problem has important research value.In addition,lightweight object detection algorithm has become one of the hot topics in the field of deep learning in recent years.High-precision convolutional neural networks often contain tens of millions of parameters and the amount of calculation,billions of high performance hardware platform is needed to achieve near real time detection speed,at cost,limited application scenarios,such as volume,mobility,therefore to lightweight improvement of network,network on resource-constrained platforms such as mobile devices and to perform real time detection with high accuracy is of important practical value.The lightweight of convolutional neural network was first investigated in this paper,the lightweight network structure SNUV was proposed,and the lightweight target detection network was designed according to structure SNUV.Channel convolution is adopted in SNUV to greatly reduce the number of parameters and the amount of computation.Meanwhile,Channel Shuffle is used to fuse the characteristics of each channel without increasing the number of parameters and the amount of computation.Compared with the original Shuffle Net V2 structure,SNUV improves the utilization of features while keeping the computational complexity unchanged.Compared with standard convolution,SNUV allows the network to have greater width and depth,which improves the accuracy and has less computational complexity.Experimental results show that snuv-based classification network and target detection network can achieve higher accuracy with less parameters and computational complexity than the same type of network.The detection of similar small targets was also studied in this paper.It is difficult to distinguish between similar targets in details,and it needs to rely on the environmental information of context to assist the detection.Therefore,sensory field is the key to distinguish similar targets.This paper puts forward a new multi-scale fusion strategy,by increasing the number of sampling reduction and fusion more receptive field characteristic figure,enables the network to better use context information such as the relative position relations,in order to make target appearance characteristics to distinguish the degree is not high up.In order to further improve the accuracy,the loss function is also improved in this paper,and the type imbalance problem of the model is improved by using Focal Loss.By introducing GIo U Loss into the loss function,the position accuracy of the model is improved.The experimental results show that the proposed multi-scale fusion strategy can effectively improve the precision of the network in the detection of similar small targets,and the improvement of the loss function further improves the network precision.In addition,the application scenarios of the algorithm was also explored in this paper,and the AR auxiliary bolt assembly system registration method based on the algorithm in this paper was also proposed.Registration is the core technology in AR system,which determines the accuracy of the superposition of virtual information in the real world and the smoothness of the picture.This system uses the algorithm in this paper as the registration method,and solves the problem of small data set size through the data augmentation strategy,thus improving the registration accuracy of the system.Detection speed and image transmission speed are the two main factors that affect the operating speed of the system.Compared with other registration methods,the algorithm in this paper has higher speed,and the transmission speed is also improved by compressing the image size.The experimental results show that the AR auxiliary bolt assembly system based on the proposed algorithm has high precision and running speed.