Research On Target Detection Algorithm Based On Convolutional Neural Network

As an important research topic in the field of computer vision,target detection is the unification of identification and positioning tasks.The main purpose of target detection is to locate the target of interest in the image,and to give the position of the bounding box of the target,and at the same time accurately determine the category of each target and score it.In this paper,the target detection algorithm based on convolutional neural network is studied.Based on the SSD algorithm,a target detection algorithm that can learn from scratch is proposed,which can be extended to the special target detection field with relatively scarce database.The algorithm architecture of this paper is mainly based on the dense block structure improved by the deep separable convolution structure.The over-fitting is well prevented while minimizing the parameter amount of the algorithm.Aiming at how to combine the detailed semantic information of low-level convolution features and the abstract information of high-level convolution features to improve the performance of features,and improve the accuracy of target detection algorithm,this paper proposes a prediction structure that fuses the semantic information of multi-layer feature graphs.This paper completed the experiment of Soft-NMS method,and analyzed and proved that the method is not suitable for regression-based target detection algorithm,and then determined that the proposed algorithm still uses the traditional NMS method.The target detection algorithm has a depth of 116 layers and a width of 48 layers.The parameter size of the algorithm is 11.8M,which is much smaller than the widely used target detection algorithms Faster R-CNN,R-FCN and SSD.The running time is 50 ms,which is basically satisfied real-time requirements.The average accuracy rate tested on the PASCALVOC data set is 78.6%,which is better than Faster R-CNN and SSD,and the algorithm of this paper especially improves the detection accuracy of small targets.The average accuracy rate after testing with the MS COCO training set is 81.6%,which is better than the current R-FCN algorithm with the highest accuracy.Therefore,the algorithm of this paper can be regarded as a real-time target detection method that takes into account the detection accuracy and speed.This paper self-made 30,000 databases with 5 types of student behavior to verify the practicability of the algorithm.The average accuracy of the algorithm in this database is 88%,which can analyze the student behavior in the video or picture in the actual classroom scene.