Research On Small Object Detection Algorithm Based On Deep Learning And Application Of Container Seal Recognition

In recent years,object detection algorithms based on deep learning have been changing rapidly,however,conventional object detection networks mainly focus on larger scale objects and pay less attention to small object detection,so the accuracy of small object detection is usually lower than that of large scale objects,coupled with the inherent problems of small objects themselves such as less available information and insignificant features,resulting in small object detection becoming a difficult problem in the field of object detection.In real scene,small object detection plays an important role in UAV object detection,remote sensing image detection,and defect detection,so it is crucial to improve the accuracy and speed of small object detection.In the process of container loading and unloading works,the container seal checking operation link still uses the manual inspection method,which has problems such as low efficiency,high labor cost and high security risk.Since the scale of lead seals is extremely small and the features are not obvious,the automatic recognition of small objects cannot be achieved by using conventional object detection networks.Based on the deep learning object detection network,this dissertation studies the current small object detection algorithm,an autonomous optimized small object detection network is proposed and applied to the container lead seal detection task to achieve fast and accurate automatic lead seal detection.Through practical application,the algorithm researched in this dissertation meets the actual engineering requirements and can replace the manual execution of the seal inspection task.The main work as well as the innovation points of this dissertation are:(1)Optimization and adjustment of feature extraction network based on Center Net object detection network.We compares Convolutional neural networks with different topologies structures used for object detection,and two types of networks with typical structures,Shuffle Net and HRNet,are studied and compared to select suitable network structures;reduce the number of network downsampling,balance the network structure parameters,improve the output feature map resolution,and improve the small object feature extraction capability;design a lightweight multi-scale feature fusion module that reduce the loss caused to small object features during network inference and improve the utilization of low-level features;adopt lightweight structures such as depth-separable convolution and Ghost Block to strike a balance between detection accuracy and detection speed.(2)Design plug-and-play small object feature enhancement module using small object context information and attention mechanism.The SSH context module is improved to enhance the small object feature saliency by extracting more effective small object context features;the three attention modules,SE module,CBAM module,and CA module,are studied and compared,and suitable attention modules are used to suppress the background noise introduced to small objects by the network inference process and the multi-scale feature fusion process to enhance small object features.(3)Optimize the detection strategy of small object detection network to improve the detection accuracy of small objects.To address the problem of imbalance between positive and negative samples for small object detection,an aspect ratio sensitive Gaussian kernel label is proposed to improve the number of positive samples and detection effect;the object detection head is redesigned to realize the interaction between the two tasks of object classification and bounding box regression,and CIOU is adopted as the evaluation index of bounding box quality to improve the small object recall rate;the use of GFocal Loss(Generalized Focal Loss)is used to further improve the effect of bounding box detection and obtain higher accuracy of small object detection,while the network can also obtain faster convergence speed,solving the problem of slow convergence of key point detection based network.(4)The improved small object detection network is applied to the container lead seal detection task.A objected data enhancement method is designed for the lead seal detection task to achieve the expansion of the number of small objects with less computation,improve the robustness of the detection network in terms of image color,object scale difference,random object location,etc.,and improve the detection effect of small objects;two detection strategies,random center clipping and image segmentation,are proposed for high resolution images,which meet different application scenarios and achieve fast detection of high resolution images.