A method and system for semantic segmentation of small objects in complex scene images

Abstract

The present invention provides a method and system for semantic segmentation of small targets in complex scene images, and relates to the field of image processing technology. The method for semantic segmentation of small targets in complex scene images includes the following contents, and establishes a feature library: selects to identify pictures based on a deep learning algorithm and stores feature information; sets segmentation parameters: the pixel size of the picture and the input minimum segmentation unit; circles and cuts the recognition area: takes the pixel size of the entire image as the segmentation unit, and takes the coincidence of the pixel origin and the segmentation unit origin as the starting point, circles and cuts in the X direction and the Y direction, and transforms the segmentation unit size until the circle cutting is completed; compares and identifies: after each circle cutting, image recognition is performed; and segmented images: the recognized target image is saved separately. Circle and select step by step for image recognition, and fully identify the target in the picture.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a method and system for semantic segmentation of small objects in complex scene images. Background Art

[0002] Image semantic segmentation is one of the three major tasks in computer vision. Its goal is to label each pixel in an image and obtain a semantic segmentation map of the image. From the perspective of traditional image segmentation, image semantic segmentation is to divide an image into multiple regions at the semantic level and then assign appropriate category labels to each region. With the development of high-speed computing devices such as GPUs, the mainstream framework for solving semantic segmentation problems has been replaced by deep learning methods.

[0003] In the existing image recognition process, there is a problem of missing small target recognition. When performing image recognition, the image is usually treated as an object for recognition, and only one feature information is obtained for this recognition, so it is impossible to effectively and accurately recognize other targets in the image. Summary of the invention

[0004] 1. Technical issues to be resolved

[0005] In view of the deficiencies in the prior art, the present invention provides a method and system for semantic segmentation of small objects in complex scene images, which solves the problem of difficulty in identifying small objects in existing image recognition.

[0006] (II) Technical solution

[0007] To achieve the above objectives, the present invention is implemented by the following technical solutions: A method for semantic segmentation of small objects in complex scene images, comprising the following contents:

[0008] Establish a feature library: select pictures of the same scene in different sizes, identify the pictures based on the deep learning algorithm, and store the feature information;

[0009] Set segmentation parameters: import the target image, obtain the image size (X, Y), and input the minimum segmentation unit (x, y);

[0010] Circle recognition area: The pixel size of the entire image is used as the segmentation unit, and the pixel origin and the segmentation unit origin coincide as the starting point. If the current segmentation unit x 当 ≥X, then line break recognition is performed. If not satisfied, then the X direction is shifted to continue the circle cutting;

[0011] The line breaks are shifted in the Y direction, and after shifting once in the Y direction, the X direction shifting circle cutting is continued, and the above process is repeated until x 当 ≥X,y 当≥Y, transform the size of the segmentation unit;

[0012] That is, by reducing the current segmentation unit (x 当 ,y 当 ) size, repeat the above process until x 当 <x, y 当 <y ends the circle cutting;

[0013] Comparison and recognition: After each encirclement, the encircled area is subjected to image recognition with reference to the feature library, and contour information is obtained;

[0014] Segment the image: Move the obtained contour outward by one pixel to obtain a new contour edge, map the contour edge coordinates to the target image, and save the image within the contour separately.

[0015] Preferably, in the encircled and identified area, each time an area is encircled, it is necessary to determine the degree of overlap between the current encircled area and the contour of the identified object. If the degree of overlap is greater than a set threshold, the current encircled area is discarded.

[0016] Preferably, the overlap degree is obtained by dividing the number of overlapped pixels between the current encircled area and the contour of the identified object by the number of all pixels in the current encircled area.

[0017] Preferably, in the encirclement identification area, the value of the displacement in the X direction is X 当 / 2-1, the displacement in the Y direction is y 当 / 2-1.

[0018] Preferably, in the encircled identification area, the segmentation unit size transformation method is (x 当 -x / 2+1,y 当 -y / 2+1).

[0019] Preferably, in the segmented image, contour parameters are saved, and a parent-child relationship is established between the saved image and the original image through the contour parameters.

[0020] A system for semantic segmentation of small objects in complex scene images, including the program written in the above and the hardware environment for running the program.

[0021] (III) Beneficial effects

[0022] The present invention provides a method and system for semantic segmentation of small objects in complex scene images. It has the following beneficial effects:

[0023] 1. The present invention circles a picture, and the circled area is from large to small, and the circled image is recognized, thereby avoiding the loss of recognition of small targets, and large targets in the picture will also appear.

[0024] 2. The present invention first determines the validity of the current area before performing image recognition. If the current area is invalid, there is no need to perform circle cutting and image recognition, which can greatly save computing resources and improve image processing efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] Example:

[0028] like Figure 1 As shown, an embodiment of the present invention provides a method for semantic segmentation of small objects in complex scene images, including the following contents:

[0029] Establish a feature library: select pictures of the same scene of different sizes, and use the pictures of different sizes to unify the features of the same target at different sizes. Use deep learning algorithms to identify the pictures and store the feature information. When identifying, it is usually done by directly identifying the obvious targets in the image, while the smaller targets are usually ignored. In addition, in traditional training, the same type of pictures but different sizes of targets are used, and their feature information is also different, so it is impossible to identify targets of all sizes in the picture.

[0030] Set segmentation parameters: import the target image and obtain the image size (X, Y), that is, construct a plane coordinate system with pixels in two directions, and input the minimum segmentation unit (x, y). The minimum segmentation unit is the size of the area that can be recognized at this size, which can be set manually.

[0031] Circle recognition area: The pixel size of the entire image is used as the segmentation unit, and the pixel origin and the segmentation unit origin coincide as the starting point. If the current segmentation unit x 当 ≥X, then line break recognition is performed. If not satisfied, then the X direction is shifted to continue the circle cutting;

[0032] The line breaks are shifted in the Y direction, and after shifting once in the Y direction, the X direction shifting circle cutting is continued, and the above process is repeated until x 当 ≥X,y 当 ≥Y, transform the size of the segmentation unit;

[0033] That is, by reducing the current segmentation unit (x 当 ,y 当 ) size, repeat the above process until x 当 <x, y 当 <y ends the circle cutting;

[0034] In the above, the size of the entire image is first used as the starting size of the segmentation unit, that is, the target on the image is first identified, and then the size of the segmentation unit is changed to segment and identify the target row by row and column by column. The process ends when the size of the segmentation unit is smaller than the set minimum segmentation unit. This method not only segments the image in terms of size, but also identifies the segmented images separately, thereby avoiding the omission of small target recognition and improving the recognition accuracy.

[0035] Comparison and recognition: After each encirclement, the encircled area is subjected to image recognition with reference to the feature library, and contour information is obtained. The image recognition and acquisition of the target contour are both carried out using existing image recognition technology;

[0036] Segment image: Move the obtained contour outward by one pixel to obtain a new contour edge, that is, increase the contour coverage range, and map the contour edge coordinates to the target image, that is, determine the position of the contour in the target image, and save the image within the contour separately. When calling to display the segmented image, the contour is highlighted in the target image to identify the position of the segmented image in the target image.

[0037] In the circled recognition area, each time the area is circled, it is necessary to determine the overlap between the current circled area and the contour of the identified object. If the overlap is greater than the set threshold, the current circled area is discarded, that is, the current circled area overlaps with the identified target. According to the overlap, the validity of the circled area at the current position is determined. If it is invalid, there is no need to perform circle cutting and image recognition, and the next step of circle cutting is performed. This step can greatly save computing resources and greatly improve the speed of image processing.

[0038] The overlap degree is calculated by dividing the number of overlapping pixels between the current circled area and the outline of the identified object by the number of all pixels in the current circled area. That is, the validity of the circled area is judged by comparing the proportion of overlapping pixels with the set threshold.

[0039] In the circle recognition area, the displacement value in the X direction is X 当 / 2-1, the displacement in the Y direction is y 当 / 2-1, that is, taking the current circle size as the boundary, move the circle row by row and column by column.

[0040] In the circle recognition area, the segmentation unit size transformation method is (x 当 -x / 2+1,y 当-y / 2+1), with the smallest segmentation unit as the size change of the segmentation unit.

[0041] In the segmented image, the contour parameters are saved, and the saved image and the original image are connected to each other through the contour parameters, that is, the saved image and the original image are connected.

[0042] In the above, segmentation is performed from region to small region, and the overlap between the segmented region and the target is discarded in the process, and then image recognition is performed. Compared with directly segmenting the image into several minimum segmented regions and then performing image recognition in sequence, it requires a lot of computing resources and there is a possibility of missing the recognition of large targets.

[0043] A system for semantic segmentation of small objects in complex scene images includes the program written in the above and the hardware environment for running the program, the hardware includes a computing unit for calculation, the computing unit is composed of a CPU and a GPU, and a storage device, etc.

[0044] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for semantic segmentation of small objects in complex scene images, characterized in that: Includes the following: Establish a feature library: select pictures of the same scene in different sizes, identify the pictures based on the deep learning algorithm, and store the feature information; Set segmentation parameters: import the target image, obtain the pixel size (X, Y) of the image, and input the minimum segmentation unit (x, y); Circle recognition area: The pixel size of the entire image is used as the segmentation unit, and the pixel origin and the segmentation unit origin coincide as the starting point. If the current segmentation unit x 当 ≥X, then line break recognition is performed. If not satisfied, then the X direction is shifted to continue the circle cutting; The line breaks are shifted in the Y direction, and after shifting once in the Y direction, the X direction shifting circle cutting is continued, and the above process is repeated until x 当 ≥X,y 当 ≥Y, transform the size of the segmentation unit; That is, by reducing the current segmentation unit (x 当 ,y 当 ) size, repeat the above process until x 当 <x, y 当 <y ends the circle cutting; Comparison and recognition: After each encirclement, the encircled area is subjected to image recognition with reference to the feature library, and contour information is obtained; Segment the image: Move the obtained contour outward by one pixel to obtain a new contour edge, map the contour edge coordinates to the target image, and save the image within the contour separately.

2. The method for semantic segmentation of small objects in complex scene images according to claim 1, characterized in that: In the encircled and identified area, each time an area is encircled, it is necessary to determine the degree of overlap between the current encircled area and the contour of the identified object. If the degree of overlap is greater than a set threshold, the current encircled area is discarded.

3. The method for semantic segmentation of small objects in complex scene images according to claim 2, characterized in that: The overlap degree is obtained by dividing the number of overlapped pixels between the current encircled area and the outline of the identified object by the number of all pixels in the current encircled area.

4. The method for semantic segmentation of small objects in complex scene images according to claim 1, characterized in that: In the circle identification area, the value of the displacement in the X direction is X 当 / 2-1, the displacement in the Y direction is y 当 / 2-1.

5. The method for semantic segmentation of small objects in complex scene images according to claim 1, characterized in that: In the circle identification area, the segmentation unit size transformation method is (x 当 -x / 2+1,y 当 -y / 2+1).

6. The method for semantic segmentation of small objects in complex scene images according to claim 1, characterized in that: In the segmented image, contour parameters are saved, and a parent-child relationship is established between the saved image and the original image through the contour parameters.

7. A system for semantic segmentation of small objects in complex scene images, characterized in that: The invention comprises a program written according to any one of claims 1 to 6 and a hardware environment for running the program.

Images