Research And Implementation Of Deep-Learning-Based Object Detection Optimizing Technology

Object detection is a basic research point in machine vision,which solves the problem of "what" and "where" in robotic environment sensing system.Specifically,object detection has two basic tasks for a given image: The first one is object classification,which gives a specific category of objects contained in the image;the second one is object location,which marks the objects with rectangular boxes(called bounding box).Real-life applications have high requirements for the accuracy and speed of object detection,such as assisted driving,artificial medical,robot applications,etc.Now,the most of object detection algorithms are based on convolutional neural networks.Although the complex network structure provides rich feature information for object classification and object localization,the huge amount of parameters of the network also makes it difficult to apply to the daily hardware equipment with poor computing ability.We optimize object detection algorithm by improving the localization accuracy and running speed to improve its usability.Most of the deep learning based optimization algorithms for improving the location accuracy of object detection are realized by deepening the network or modifying the network structure.A more sophisticated network can not only increase the detection accuracy,but also reduce the running speed.To solve this problem,this paper proposes a border-oriented post-processing optimization algorithm called BOPP(Border-Oriented Post-Processing),which optimizes localization accuracy of a detected bounding box by refining the four borders.Instead of modifying the existing object detection algorithm,BOPP realizes an independent optimization module to optimize the localization accuracy of the detected bounding box.This independent design pattern allows it to be used to optimize any of the existing object detection algorithms.In addition,unlike the algorithms based on convolutional neural networks,BOPP uses HOG features and random forests as the basis of the algorithm,so its features and regressors are strongly interpretable.Both feature extraction and regression can in high speed.Experiments show that BOPP can get nearly 8% mAP improvements of YOLO and Faster R-CNN algorithms over the IoU threshold of 0.8 on the KITTI dataset.The time for optimizing one bounding box on the CPU is about 9 milliseconds.Fast and efficient optimization as well as good scalability make BOPP ideal for optimizing the algorithms with high speed but poor accuracy,such as YOLO.To improve the running speed,this paper combines the image processing network MobileNet with the object detection algorithm YOLO,delete some layers of the network and construct a lightweight deep learning based object detection algorithm,named Mobile-YOLO.Specifically,we have simplified YOLO's network structure and converted the convolution into a depthwise / pointwise pattern convolution,and then we also do a series of network compression processes.Comparing with conventional convolutional YOLOs,the parameters of Mobile-YOLO are compressed by one hundred times.A smaller amount of parameters will inevitably lead to a reduction in computation and an increase in network model running speed.Experiments show that,in the same configuration of CPU computing nodes,Mobile-YOLO has more than 5 times the speed compared to the original YOLO algorithm.In addition,in order to compensate for the loss of accuracy,we introduced the two-class classification into Mobile-YOLO,which improved the discrimination ability of the model for similar categories and reduced the number of misclassifications,thus improving the accuracy of object detection.In other words,Mobile-YOLO can run very fast with comparable accuracy with the original YOLO algorithm.Finally,to further improve the speed of the model,we sacrificed some of the precision.By cutting the network layers of Mobile-YOLO,we got a faster model named Fast-Mobile-YOLO,which runs in real-time on the CPU with the speed of 23 fps.