Small target detection method and apparatus in image based on depth learning

Abstract

Embodiments of the present invention provide a small target detection method and apparatus in an image based on depth learning. The method comprises: obtaining a to-be-detected image; obtaining a target category of the to-be-detected image and location coordinates of the target category in the to-be-detected image based on the to-be-detected image and a pre-trained target detector model, wherein the process comprises: inputting the to-be-detected image into a target feature extractor to obtain a feature map, inputting the feature map into a target area to generate a network to obtain coordinates of a candidate box, inputting the coordinates of the candidate box into a context information layer, calculating by a preset calculation manner according to the coordinates of the candidate box toobtain coordinates of the vertical candidate box and coordinates of the horizontal candidate box; and inputting the coordinates of each candidate box and the feature map into a target area classification network to obtain the category and location coordinates of the target. According to the technical scheme of the embodiments of the present invention, even for the relatively small target in the image, more feature information is obtained due to the target area classification network, so that the accuracy when detecting the small target such as the traffic sign is improved.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a method and device for detecting a small target in an image based on deep learning. Background technique [0002] The detection of targets in images is widely used in various fields. For example, in the field of automatic driving, the detection of traffic signs in images is a very important link. Identify and guide the driving of vehicles to ensure driving safety. [0003] Among the target detection techniques in images, the Faster r-cnn (accelerated regional convolutional neural network) detector is most commonly used. The detector consists of three parts, the feature extractor, the region generation network and the region classification network, and the Faster r-cnn detector needs to be trained before the actual detection. In the actual detection process, the image is first input into the feature extractor for convolution operation to obtain the feature map of the e...

AI Technical Summary

Problems solved by technology

Even if the region generation network accurately provides the location of the target, because the feature information of the target is less than one pixel, the feature information is too scarce, and the region classification network cannot be classified.

Images