Method for realizing tiny target detection on chip integrating distillation strategy and deconvolution

Abstract

A method for realizing tiny target detection on a chip integrating a distillation strategy and deconvolution trains a plurality of layers in a learning network for a low-resolution image and containing a deconvolution layer in a loss-resistant learning mode by a plurality of intermediate feature maps of a teaching network for a high-resolution image, so that the output precision of the learning network is improved and the size of the chip is reduced while the receptive field of a low-pixel image is enlarged. The method simply designs the target detection task as a classification task through the learning network, and only needs to determine whether the target exists in an area of 20*20 pixels, so that the on-chip area is small when the hardware is realized, the required memory is small, and meanwhile, the false detection of the tiny objects can be effectively eliminated.

Description

technical field [0001] The present invention relates to a technology in the field of image detection, in particular to an on-chip implementation method of tiny target detection integrating a distillation strategy and deconvolution. Background technique [0002] Although there are currently several methods for object detection using convolutional neural networks, most popular algorithms perform well when objects occupy a large portion of an image (usually their size is greater than 20 square pixels). Recently, many algorithms have emerged to detect tiny objects with low resolution (less than 20 square pixels). These methods usually rely on multi-scale resolution and detect objects of different sizes at corresponding resolutions. This architecture detects multiple target objects at the same time (multitasking), which is beneficial to improve the detection of tiny target objects, but requires larger storage capacity and longer computing time in the hardware implementation proc...

AI Technical Summary

Problems solved by technology

This architecture detects multiple target objects at the same time (multitasking), which is beneficial to improve the detection of tiny target objects, but requires larger storage capacity and longer computing time in the hardware implementation process

Images