CN112183579B - Method, medium and system for detecting micro target - Google Patents

Abstract

The invention discloses a micro target detection method, medium and system. The method comprises: after inputting an image containing a tiny target into a deep learning target detection network, outputting the tiny target extracted from the image, wherein the area of the tiny target is smaller than a preset area; The tiny target is input into the amplification deep learning network, and the tiny target with a preset magnification is output; the tiny target with a preset magnification is input into the deep learning classification network, and the category of the tiny target is output. The invention can effectively improve the detection accuracy of tiny targets, reduce the false detection rate, thereby improving the performance and stability of the entire computer vision system, and has good practical application value and economic benefits.

Description

A tiny target detection method, medium and system

technical field

The invention relates to the technical field of target detection, in particular to a tiny target detection method, medium and system.

Background technique

In recent years, with the continuous development and breakthrough of deep learning and computer computing power, deep learning has achieved incomparable effects in traditional algorithms in computer vision fields such as classification tasks, object detection, and semantic segmentation, and has been applied in various industries.

In the field of target detection, the current mainstream application algorithms are: SSD, FasterRcnn, YOLO, etc., which have achieved good results in practical applications, but a common problem with these algorithms is that when detecting tiny targets, their detection accuracy is low. . The main reason is firstly that tiny targets contain fewer pixels and carry less data information, and secondly, the general rule of deep learning is that the deeper the network depth, the higher the detection accuracy. This will cause the tiny target to pass through a deep deep learning network. Under the effect of downsampling, the less information carried by the tiny target itself will almost disappear in the deeper feature map, resulting in a decrease in the detection accuracy of the tiny target.

Contents of the invention

Embodiments of the present invention provide a micro-target detection method, medium and system to solve the problem of low precision in detecting micro-targets in the prior art.

In the first aspect, a method for detecting a tiny target is provided, comprising: after inputting an image containing a tiny target into a deep learning target detection network, outputting the tiny target extracted from the image, wherein the area of the tiny target is Smaller than the preset area; input the extracted tiny target into the enlarged deep learning network, and output the tiny target enlarged by a preset multiple; input the tiny target enlarged by a preset multiple into the deep learning classification network, and output the tiny target The category of the target.

In the second aspect, a computer-readable storage medium is provided, and computer program instructions are stored on the computer-readable storage medium; when the computer program instructions are executed by a processor, the tiny target detection as described in the embodiment of the first aspect is realized method.

In a third aspect, a tiny target detection system is provided, including: the computer-readable storage medium as described in the embodiment of the second aspect.

In this way, the embodiment of the present invention can effectively improve the detection accuracy of tiny targets, reduce the false detection rate, thereby improving the performance and stability of the entire computer vision system, and has good practical application value and economic benefits.

Description of drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings that need to be used in the description of the embodiments of the present invention. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention , for those skilled in the art, other drawings can also be obtained according to these drawings without paying creative labor.

Fig. 1 is the flow chart of the tiny object detection method of the embodiment of the present invention;

Fig. 2 is a schematic structural diagram of an enlarged deep learning network according to an embodiment of the present invention;

3 is a schematic diagram of a local structure of a residual ResNet network according to an embodiment of the present invention;

Fig. 4 is a schematic diagram of a periodic card mixing operator according to an embodiment of the present invention.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some of the embodiments of the present invention, but not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

The embodiment of the invention discloses a tiny target detection method. As shown in Figure 1, the tiny target detection method includes the following steps:

Step S1: After inputting the image containing the tiny target into the deep learning target detection network, output the tiny target extracted from the image.

The deep learning target detection network uses a well-known neural network, for example, YoloV3 target detection network, SSD target detection network, FasterRcnn target detection network. It should be understood that, generally, the images input to the deep learning object detection network should be scaled to a certain size. Therefore, before step S1, the method in the embodiment of the present invention further includes: scaling the original image into an image with preset pixels. The preset pixels can be determined according to the selected deep learning target detection network. For example, for the YoloV3 object detection network, the scaled image is a 608×608 (px) image. Specifically, the original image can be read by the imread method of opencv, and then the read original image can be scaled to a size of 608×608 by the resize method of opencv.

The area of the tiny target described in the embodiment of the present invention is smaller than a preset area. For example, in the embodiment of the present invention, the preset area is 20×20 (px). It should be understood that the tiny target is for the image input to the deep learning target detection network, not the original image.

Specifically, the deep learning target detection network can output the width and height of the image, and the area can be calculated according to the product of the width and height, so as to determine whether it belongs to a tiny target.

Step S2: Input the extracted tiny target into the amplified deep learning network, and output the tiny target with a preset magnification factor.

Specifically, as shown in Figure 2, the structure of the enlarged deep learning network consists of sequentially connected micro-target ResNet network structural blocks and sub-pixel convolutional layers.

Among them, the micro target ResNet network structure block is composed of the residual ResNet network with a preset number of cascades, forming a deep network structure, which can fully extract the deep semantic information of the micro target. In the embodiment of the present invention, the residual ResNet network is cascaded 9 times. As shown in Figure 3, the two-layer weight layer (weight layer) in the residual ResNet network is replaced by a dense DenseNet network. The residual ResNet network and the dense DenseNet network are well-known network structures, and will not be described here.

The output result of the residual ResNet network X _l =H _l (X _l-1 , _wi , _bi )+X _l-1 . Among them, X _l-1 represents the input object of the current residual ResNet network. H _l represents a function of the dense DenseNet network. w _i and b _i represent the parameters of the dense DenseNet network, which can be obtained by model training in combination with examples.

The tiny target feature map F _l-1 =R(X _lr ,W,B) output by the tiny target ResNet network structure block. Among them, X _lr represents the tiny target input to the building block of the tiny target ResNet network. R represents the nonlinear function of the small target ResNet network structure block, generally a Relu nonlinear function. W and B represent the parameter weights and biases of the micro-target ResNet network building block, which can be obtained by model training in combination with examples. The size of the tiny object feature map is w×h×c×r ² . r represents the magnification of the tiny object feature map. w×h×c represents the size of the tiny target input into the tiny target ResNet network building block, w represents the width of the tiny target input into the tiny target ResNet network building block, h represents the height of the tiny target input into the tiny target ResNet network building block, c Indicates the number of image channels, generally 3, namely RGB three channels.

By combining the residual ResNet network and the dense DenseNet network, a new network structure block is proposed. The residual ResNet network ensures the depth of the network, and the dense DenseNet network ensures that more tiny target information is transferred to the deep feature map.

Since the tiny target itself contains less information due to fewer pixels, in order to make the enlarged target carry more target feature information, the embodiment of the present invention uses sub-pixel convolution to enlarge the tiny target. Specifically, the sub-pixel convolution formula of the sub-pixel convolution layer is I _sr =PS(W _l ×F _l-1 +B _l ). Among them, I _sr represents the output result of the sub-pixel convolutional layer. F _l-1 represents the tiny target feature map output by the tiny target ResNet network building block. PS represents the periodic card mixing operator. As shown in Figure 4, the low-pixel feature map is obtained by r ² convolution kernels to obtain r ² feature maps, and then the sub-pixel points in these r ² feature maps are divided from left to right. To the right, arranged in order from top to bottom to get the final enlarged image. W _l and B _l represent sub-pixel convolution weights and biases respectively, which can be obtained by model training in combination with examples. In a specific embodiment of the present invention, the preset multiplier is 8 times, and the output micro target is an image of 8*w _i ×8*h _i ×3, where 3 represents three channels of R, G, and B.

Step S3: input the tiny target enlarged by a preset multiple into the deep learning classification network, and output the category of the tiny target.

The deep learning classification network adopts a well-known neural network, for example, the inceptionv4 classification network, the vgg16 classification network, and the resnet50 classification network.

In a specific embodiment of the present invention, the inceptionv4 classification network is used to output the result (p _i , ci ₎ . Among them, p _i represents the confidence of the i-th tiny target, and c _i represents the classification result of the i-th tiny target. Generally, the classification result is determined by specific examples, for example, pedestrians and vehicles may be included.

In addition to the classification results of the above-mentioned tiny targets, the method of the embodiment of the present invention can also detect the location information of the tiny targets. Preferably, the tiny target detection method of the embodiment of the present invention further includes the following steps:

After the image containing the tiny target is input into the deep learning target detection network, the location information of the tiny target is output.

The deep learning object detection network is as described in the aforementioned step S1, and will not be repeated here. Output the location information (xi _{, y ,} w, h _i ) of all tiny targets through the deep learning target detection network. Among them, x _i represents the abscissa of the center of the i-th micro-target, y _i represents the ordinate of the center of the i-th micro-target, w _i is the width of the i-th micro-target, h _i is the width of the i-th micro-target high. w _i and _hi can be used to calculate the area of the output target, so as to determine whether the target is a tiny target.

Therefore, the tiny target detection method of the embodiment of the present invention can not only detect the category of the tiny target, but also obtain the location information of the tiny target, which can be used (p _i , c _i , x _i , y _i , w _i , h _i ) Represents the comprehensive result including tiny object category and location information.

The embodiment of the present invention also discloses a computer-readable storage medium, where computer program instructions are stored on the computer-readable storage medium; when the computer program instructions are executed by a processor, the tiny target detection as described in the above-mentioned embodiments is realized method.

The embodiment of the present invention also discloses a tiny target detection system, including: the computer-readable storage medium as described in the above-mentioned embodiments.

To sum up, the embodiments of the present invention can effectively improve the detection accuracy of tiny targets, reduce the false detection rate, thereby improving the performance and stability of the entire computer vision system, and have good practical application value and economic benefits.

The above is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Anyone skilled in the art can easily think of changes or substitutions within the technical scope disclosed in the present invention. Should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims (9)

1. A tiny target detection method, characterized in that, comprising: After inputting the image containing the tiny target into the deep learning target detection network, outputting the tiny target extracted from the image, wherein the area of the tiny target is smaller than a preset area; Inputting the extracted tiny target into the amplified deep learning network, and outputting the tiny target with a preset multiplier; Inputting the micro-target enlarged by a preset multiple into a deep learning classification network, and outputting the category of the micro-target; The structure of the enlarged deep learning network is composed of sequentially connected micro target ResNet network structure blocks and sub-pixel convolution layers, wherein the micro target ResNet network structure blocks are composed of residual ResNet networks cascaded for preset times, and the The two-layer weight layer in the residual ResNet network is replaced by a dense DenseNet network, the residual ResNet network ensures the depth of the network, and the dense DenseNet network ensures that more tiny target information is transferred to the deep feature map. 2. The tiny target detection method according to claim 1, characterized in that: the output result X _l of the residual ResNet network = H _l (X _l-1 , w _i , b _i )+X _l-1 , Among them, X _1-1 represents the input object of the current residual ResNet network, H ₁ represents the function of the dense DenseNet network, and w _i and _bi represent the parameters of the dense DenseNet network. 3. The tiny target detection method according to claim 1, characterized in that: the tiny target feature map F _l-1 =R(X _lr , W, B) output by said tiny target ResNet network structure block, wherein, X _lr represents the tiny target that is input into the tiny target ResNet network structure block, R represents the nonlinear function of the tiny target ResNet network structure block, W and B represent the parameter weights and biases of the tiny target ResNet network structure block value. 4. The tiny target detection method according to claim 3, characterized in that: the size of the tiny target feature map is w×h×c×r ² , r represents the magnification of the tiny target feature map, w× h×c represents the size of the tiny object input to the tiny object ResNet network building block. 5. The tiny target detection method according to claim 1, characterized in that: the sub-pixel convolution formula of the sub-pixel convolution layer is I _sr =PS(W _l ×F _l-1 +B _l ), where , I _sr represents the output result of the sub-pixel convolutional layer, F _l-1 represents the micro target feature map output by the micro target ResNet network structure block, PS represents the periodic mixed card operator, W _l and B _l represent Subpixel convolution weights and biases. 6. The tiny target detection method according to claim 1, further comprising: After the image containing the tiny target is input into the deep learning target detection network, the location information of the tiny target is output. 7. The tiny target detection method according to claim 1 or 6, wherein before the step of inputting the image containing the tiny target into the deep learning target detection network, the method further comprises: Scales the original image to said image at preset pixels. 8. A computer-readable storage medium, characterized in that: computer program instructions are stored on the computer-readable storage medium; when the computer program instructions are executed by a processor, the computer program instructions described in any one of claims 1 to 7 are implemented. The micro-target detection method described above. 9. A tiny target detection system, comprising: the computer-readable storage medium according to claim 8.

Images