A high-precision and fast single-class object detection method based on deep learning

Abstract

A high-precision and fast single-class target detection method based on deep learning, including the following steps: Step 1) Designing a neural network: determining the characteristics of the target, the characteristics include: the target is a rigid body or a non-rigid body, and the target is self-contained Or make a part of a larger object; determine the size change range of the target object in the application scene, determine the range that the neuron receptive field needs to cover under different sizes according to the characteristics of the target object and the size change range of the target object, and then design A neural network structure of a specific depth. Step 2) Neural network training: collect the image data containing the target object in the target scene and mark the target object in each image, and use the marked image data as the target of network learning; execute the standard neural network training process ; Step 3) After the neural network is trained, the test scene image is input to the neural network, and the target can be located according to the output of the neural network.

Description

technical field [0001] The invention relates to technical fields such as artificial neural networks and computer vision, and in particular to a high-precision and fast single-class target detection method based on deep learning. Background technique [0002] Target detection is one of the basic problems in the field of computer vision. The task it needs to complete is to locate multiple specific categories of targets (such as people, cars, trees, etc.) ) marked. Specifically, the present invention relates to single-category target detection, that is, there is only one type of target to be located and marked. Multi-class object detection covers the single-class case. [0003] In the era before deep learning, the classic framework to solve the problem of object detection is a sliding window combined with artificial image feature classification. For example, face detection usually uses LBP (Local Binary Pattern) features plus Adaboost classifier; pedestrian detection usually...

AI Technical Summary

Problems solved by technology

However, the current second type of method predefines a number of anchor boxes for the neurons in the upper layer of the neural network as the target candidate area, and at the same time corrects the position of the anchor box, the amount of calculation is still large, and there is room for further speed-up

Images