Disclosure of Invention
In order to overcome the defects in the prior art, the embodiment of the invention provides a small target object detection method, which comprises the following steps:
s1, continuously and repeatedly carrying out down-sampling on the original image where the small target object is located by using the convolutional neural network model to obtain a plurality of characteristic images;
s2, the convolutional neural network model sorts the plurality of feature images according to the time sequence of the down sampling to obtain a feature image group;
s3 selecting the last n characteristic images in the characteristic image group by the convolutional neural network model, and sequentially marking the n characteristic images as C according to the size sequence1、C2、C3…Cn;
S4 characteristic image C is processed by using convolution neural network modelnPerforming convolution operation by 1 to obtain a characteristic image Mn;
S5 performing M on the characteristic image by using the convolutional neural network modelnUp-sampling to obtain size and characteristic image C4Consistent feature image Mn -;
S6 characteristic image C is processed by using convolution neural network modeln-1Performing convolution operation by 1 to obtain a characteristic image Cn-1 -;
S7 using convolution nerve network model to analyze the characteristic image Mn -And the characteristic image Cn-1 -Overlapping to obtain a characteristic image Mn-1;
S8 repeating the above steps S6-S7 to obtain the feature image Mn-2… characteristic image M2Characteristic image M1;
S9 using convolution neural network model to respectively process the characteristic images M1Characteristic image M2Characteristic image M3… characteristic image Mn-1Characteristic image MnPerforming convolution operation to make the characteristic image M1Characteristic image M2Characteristic image M3… characteristic image Mn-1Characteristic image MnThe number of output channels is the same;
the S10 convolutional neural network model adopts the sizes of [8 × 8, 16 × 16, 32 × 32, 64 × 64, 128 × 128 … 2 [ ]n+2*2n+2]n groups of prediction frames respectively predict characteristic image M1Characteristic image M2Characteristic image M3… characteristic image Mn-1Characteristic image MnSmall target object in (1).
Preferably, the feature image M1Characteristic image M2Characteristic image M3… characteristic image Mn-1Characteristic image MnThe number of output channels is 256.
Preferably, n has a value of 5.
The small target object detection method provided by the embodiment of the invention has the following beneficial effects:
the method fully utilizes the characteristic images acquired in each convolution and downsampling process, superposes each characteristic with the characteristic output by downsampling next time, realizes the prediction of small target objects by using large-size characteristic images with less lost information, and predicts large target objects by using small-size characteristic images with more abstract semantic information obtained through multiple convolutions, so that the final prediction result is more accurate.
Detailed Description
The invention is described in detail below with reference to the figures and the embodiments.
As shown in fig. 1, the method for detecting a small target object provided by the embodiment of the present invention includes the following steps:
s101, continuously and repeatedly carrying out down-sampling on an original image where a small target object is located by using a convolutional neural network model to obtain a plurality of characteristic images;
s102, sequencing a plurality of feature images by the convolutional neural network model according to the time sequence of the down sampling to obtain a feature image group;
s103, selecting the last n characteristic images in the characteristic image group by the convolutional neural network model, and sequentially marking the n characteristic images as C according to the size sequence1、C2、C3…Cn;
S104, utilizing the convolution neural network model to perform comparison on the characteristic image CnPerforming convolution operation by 1 to obtain a characteristic image Mn;
S105, performing M on the characteristic image by using the convolutional neural network modelnUp-sampling to obtain size and characteristic image C4Consistent feature image Mn -;
S106, using the convolution neural network model to perform characteristic image Cn-1Performing convolution operation by 1 to obtain a characteristic image Cn-1 -;
S107, utilizing the convolution neural network model to carry out feature image Mn -And the characteristic image Cn-1 -Overlapping to obtain a characteristic image Mn-1The fusion of information among different characteristic images is realized through pixel-by-pixel addition;
s108, repeating the steps S106-S107 to obtain the characteristic image Mn-2… characteristic image M2Characteristic image M1;
S109, respectively carrying out comparison on the characteristic images M by using the convolutional neural network model1Characteristic image M2Characteristic image M3… characteristic image Mn-1Characteristic image MnPerforming convolution operation to make the characteristic image M1Characteristic image M2Characteristic image M3… characteristic image Mn-1Characteristic image MnThe number of output channels is the same;
s110, adopting the sizes of the convolutional neural network model as [8 × 8, 16 × 16, 32 × 32, 64 × 64, 128 × 128 … 2n +2*2n+2]n groups of prediction frames respectively predict characteristic image M1Characteristic image M2Characteristic image M3… characteristic image Mn-1Characteristic image MnSmall target object in (1).
And (4) sliding the prediction frame on the characteristic image, classifying the result after each sliding, and realizing the detection on the small target object in the characteristic image.
Optionally, the characteristic image M1Characteristic image M2Characteristic image M3… characteristic image Mn-1Characteristic image MnThe number of output channels of (2) is 256.
Wherein the number of output channels is related to the parameters and the computing power of the convolutional neural network model.
Optionally, n has a value of 5.
Generally, the larger the value of n, the more accurate the final prediction result, but since up-sampling is performed once each time, the more the calculation amount is increased later, and therefore, 5 times are values that are preferably selected after being weighted between the accuracy and the calculation amount.
According to the small target object detection method provided by the embodiment of the invention, each feature is superposed with the feature output by the next downsampling by using the feature image acquired in each convolution and downsampling process, so that the small target object is predicted by using the large-size feature image with less lost information, the large target object is predicted by using the small-size feature image with more abstract semantic information obtained through multiple convolutions, and the accuracy of the prediction result is improved.
In the foregoing embodiments, the descriptions of the respective embodiments have respective emphasis, and for parts that are not described in detail in a certain embodiment, reference may be made to related descriptions of other embodiments.
It should also be noted that the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other identical elements in the process, method, article, or apparatus that comprises the element.
As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The above are merely examples of the present application and are not intended to limit the present application. Various modifications and changes may occur to those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.