Object detection method based on deep network

Abstract

The method comprises the following steps: firstly, preprocessing data set pictures; training a deep network to obtain an optimal structure and parameters; extracting convolutional characteristics of an object through the deep network after parameter optimization, calculating similarity through the characteristic vectors, achieving object detection, and the method specifically comprises the three steps that an image data set is preprocessed to obtain training data, the training data is input to train the deep network, and the trained deep network is adopted for object detection. According to the method, the distribution rule of the original number is not lost while data normalization is carried out, the learning efficiency is greatly improved, the training speed is greatly increased, the classification effect is not affected, the accuracy of initial parameters does not need to be pursued during parameter initialization, training is rapidly converged, and therefore the effects of being rapid in training and high in detection rate are achieved.

Description

technical field [0001] The present disclosure relates to the field of object detection in image processing, to deep learning technology, and in particular to a deep network-based object detection method. Background technique [0002] With the development of intelligence, image object detection has become an important technology, which has a wide range of applications in various industries, such as machine vision, intelligent transportation, security monitoring, aerospace military and other fields. There are various methods of object detection, which can be roughly divided into two categories. One is object detection methods based on local features, such as color-based object detection methods and texture-based object detection methods. Such methods combine multiple features , the accuracy of pedestrian detection is very high, but due to the high dimensionality of the feature vector of this method, the detection speed is relatively slow. These basic features are finally clas...

AI Technical Summary

Problems solved by technology

This type of method can use the sliding window method for convolutional feature extraction, but this type of method needs to scan the entire picture, and the speed is very slow

For example, the typical deep networks of fast-rcnn and faster-rcnn, two existing technologies applied to image processing, converge only after 2000 iterations, and the time complexity is extremely high.

Images