A Faster RCNN Object Detection Method Based on Difficult Sample Mining

Abstract

The invention discloses a Faster RCNN target detection method based on refractory sample mining. The method comprises the steps of 1, image target detection based on deep learning; 2, on the basis ofan online sample mining method, adjusting and setting adopted key parameters as follows: step 3, negative difficult-to-distinguish sample mining: on the basis of a difficult-to-distinguish sample, adjusting a mini-formed by an RPN in training; wherein the ratio of the positive sample to the negative sample of the bch is 1: 3; and step 4, removing redundant frames to avoid loss of multiple calculations. An improved non-maximum suppression algorithm is adopted to reasonably remove redundancy from a suggestion box generated by an RPN layer network. Under the condition that samples are not expanded. The definition of negative samples is broadened, and more samples difficult to train are mined online through the samples. The ratio of positive samples to negative samples is set, and rare sampleswith the maximum loss and difficult training are reasonably and simply calculated; and the loss of classification and frame regression is balanced, so that the continuous reduction of training loss can be met.

Description

technical field [0001] The invention relates to a method based on the combination of online sample mining technology (OHEM) and negative hard sample mining (HNEM). [0002] technical background [0003] In recent years, with the rapid development of computer science and technology, image processing and image target detection based on computer technology have also achieved unprecedented rapid development. Among them, deep learning extracts key target features by learning massive digital image features. It has surpassed human beings, bringing one surprise after another to the industry. The two major tasks of computer vision are image classification and target location detection, while the traditional method of image feature detection relies on experienced algorithm engineers to design matching templates corresponding to targets, such as deformable model (DPM), HOG feature extraction, etc. The former's Target detection is to locate the target through a sliding window, and then ...

AI Technical Summary

Problems solved by technology

The existing problems are that the detection time is long, the feature matching accuracy is low, and it is effective for specific targets, resulting in a series of problems such as weak model generalization ability.

The recently emerging deep learning method can effectively identify features for targets in complex scenes, and the recognition effect is far superior to traditional methods, but there are also shortcomings: (1) The training set for deep learning requires a large amount of data. Thousands of samples can effectively learn data features, and the more data, the better the effect, which brings certain difficulties to the data collection work; (2) The hardware requirements are high, and training large data samples requires at least 4 G (3) The training skills are strong. When the parameter settings are unreasonable, the sample training effect is poor and difficult to train.

Images