CN115761449A - High-resolution image small target detection method - Google Patents

Abstract

The invention discloses a high-resolution image small target detection method, which adopts a blocking scheme to process aiming at high resolution, divides an original image into a plurality of subgraphs by off-line multi-scale cutting in a training stage, and trains a network by utilizing the subgraphs; in the inference stage, prediction is carried out in the form of an online multi-scale sliding window. The method is suitable for the target detection problem of high-resolution images such as aerial images, satellite remote sensing images, array images and the like.

Description

High-resolution image small target detection method

Technical Field

The invention relates to the technical field of image processing, in particular to a method for detecting a small target of a high-resolution image.

Background

The high-resolution image cannot be directly used for deep neural network input training or detection, because image information is lost due to down-sampling, the detection performance of the small target is influenced, on one hand, the down-sampling causes the small target to have the pixel size of single digit basically in the size of a characteristic diagram, and the designed target detection classifier has poor classification effect on the small target; on the other hand, the scope of the small target features mapped back to the original image due to the downsampling process may be larger than the size of the small target on the original image, which results in poor detection effect.

Disclosure of Invention

The invention aims to solve the technical problem of detecting a small target of a high-resolution image and provides a method for detecting the small target of a deep neural network based on image segmentation.

The technical scheme adopted by the invention for solving the technical problem is as follows: a high-resolution image small target detection method comprises the following steps:

s1, training process

S11, constructing an algorithm training set: cutting the original high-resolution image into pictures according to 1/4, 1/8 and 1/16, blocking according to the proportion of an overlapping area of 0.2, and generating a multi-scale blocked image offline as a training image to perform algorithm training with a target detection label file; intercepting an area image by adopting a sliding window from left to right and from top to bottom, intercepting the intersection ratio IOU of a target/a complete target to be more than 0.5, generating a new marked coordinate of the target, and storing a multi-scale block diagram and a target detection label file to obtain a new training sample;

s12, initializing a deep learning target detection model randomly, importing a training sample into the deep learning target detection model to carry out training data random sampling, predicting to obtain a detection frame pixel coordinate and detection frame classification information, calculating a regression and classification loss function and a gradient, updating algorithm model parameters, judging whether set iteration times are reached, if so, finishing training, otherwise, repeatedly carrying out training data random sampling;

s13, obtaining a trained deep learning target detection model;

s2, reasoning process

S21, a high-resolution image online sliding window reasoning algorithm: generating sliding windows of the original high-resolution image according to 1/4, 1/8, 1/16 and the proportion of the overlapping area of 0.2, and sliding the windows from left to right and from top to bottom;

s22, sending the image in the sliding window area into a deep learning target detection model to obtain an area image detection result;

s23, resolving the target pixel coordinates in the sliding window back to the pixel coordinates corresponding to the original high-resolution image, predicting to obtain the pixel coordinates of a detection frame and classification information of the detection frame, outputting the pixel coordinates and classification information of the detection target, and loading the trained model weight file to a deep learning target detection model;

and S24, inhibiting redundant detection frames by adopting a global NMS algorithm and an In-box NMS algorithm to obtain a final result.

The beneficial effects of the invention are: the detection method adopts an image blocking mode to process in an algorithm training stage, a training set is cut into small blocks of images in an off-line mode and is input into a deep neural network for training, online reasoning is carried out in a sliding window mode in a reasoning stage, the proper image blocking size and the proportion of an overlapped area between the blocks can be determined according to the size of a target in an original image, and compared with the current method of directly zooming a high-resolution image and inputting a deep learning target detection model, the detection performance of the small target in the high-resolution image is better. The method is suitable for the target detection problem of high-resolution images such as aerial images, satellite remote sensing images, array images and the like.

Drawings

FIG. 1 is an overall framework of the algorithm of the present invention;

FIG. 2 is a flow chart of a high-resolution image block training and online reasoning algorithm.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

As shown in FIG. 1, the present invention provides an algorithm overall framework of a high resolution image small target detection method based on image blocking. The whole algorithm framework consists of a training phase and an inference phase.

And processing the original high-resolution image by adopting a blocking scheme, in a training stage, intercepting the image of the area by adopting an offline multi-scale blocking scheme according to the configuration of the original resolution ratio of 1/4, 1/8 and 1/16 and the proportion of the overlapping area of 0.2 and adopting a sliding window from left to right and from top to bottom, generating a new target labeling coordinate if the intersection ratio (IOU) of the intercepted target/the complete target is greater than 0.5, and storing the multi-scale blocking diagram and the label file to obtain a new training sample. And inputting the subgraph and the label file into a deep learning target detection algorithm model for training to obtain a trained model weight file.

And training a deep learning target detection model by using the block images to obtain a trained target detection model.

In the step, an original image is split into a plurality of sub-images through off-line multi-scale cutting, the sub-images and corresponding label files are generated, and a network is trained by using the sub-images.

And in the inference stage, the original high-resolution image adopts online multi-scale sliding window inference, a sliding window is generated according to the settings of 1/4, 1/8 and 1/16 of the original resolution and the proportion of 0.2 of the overlapping area in sequence, and the sliding window slides on the high-resolution image from left to right and from top to bottom in sequence. And predicting in the form of an online multi-scale sliding window.

And loading the trained model weight file by the deep learning target detection algorithm model, and sending the image in the sliding window area into the target detection model to obtain the area image detection result.

And the sliding window slides in the original high-resolution image from left to right and from top to bottom sequentially, the image in the sliding window is input into the target detection model, and the pixel coordinate information of the detected target is output.

And resolving the coordinates of the target pixels in the sliding window back to the coordinates of the corresponding pixels of the original high-resolution image. That is, the detection result is converted back to the original image according to the pixel coordinates of the sliding window on the original image. And adopting a global NMS and In-box NMS algorithm to inhibit redundant detection frames to obtain a final detection target coordinate information result.

As shown in FIG. 2, the invention provides an algorithm training and reasoning process of a high-resolution image small target detection method based on image segmentation, and pseudo codes are described as follows.

Algorithm 1: high-resolution image blocking training algorithm

And 2, algorithm: high-resolution image online sliding window reasoning algorithm

The above-described embodiments are merely illustrative of the principles and effects of the present invention, and some embodiments may be applied, and it will be apparent to those skilled in the art that various changes and modifications may be made without departing from the inventive concept of the present invention, and these embodiments are within the scope of the present invention.

Claims (2)

1. A method for detecting a small target of a high-resolution image is characterized by comprising the following steps: comprises the following steps

S1, training process

S11, cutting the original high-resolution image into pictures according to 1/4, 1/8 and 1/16, blocking according to the overlapped area, and generating a multi-scale blocked image in an off-line manner to serve as a training image and a target detection label file for algorithm training; intercepting an area image by adopting a sliding window from left to right and from top to bottom, intercepting the intersection ratio IOU of a target/a complete target to be more than 0.5, generating a new marked coordinate of the target, and storing a multi-scale block diagram and a target detection label file to obtain a new training sample;

s12, initializing a deep learning target detection model randomly, importing a training sample into the deep learning target detection model to carry out training data random sampling, predicting to obtain a detection frame pixel coordinate and detection frame classification information, calculating a regression and classification loss function and a gradient, updating algorithm model parameters, judging whether set iteration times are reached, if so, finishing training, otherwise, repeatedly carrying out training data random sampling;

s13, obtaining a trained deep learning target detection model;

s2, reasoning process

S21, generating sliding windows of the original high-resolution image according to 1/4, 1/8, 1/16 and the overlapping area, and sliding the windows from left to right and from top to bottom;

s22, sending the image in the sliding window area into a deep learning target detection model to obtain an area image detection result;

s23, resolving the target pixel coordinates in the sliding window back to the pixel coordinates corresponding to the original high-resolution image, outputting the pixel coordinates and category information of the detection target, and loading the weight file of the trained model to the deep learning target detection model;

and S24, inhibiting redundant detection frames by adopting a global NMS algorithm and an In-box NMS algorithm to obtain a final result.

2. The method as claimed in claim 1, wherein the ratio of the overlapping area is 0.2.

Images