CN111582093A - An automatic detection method for small objects in high-resolution images based on computer vision and deep learning - Google Patents

Abstract

The invention discloses an automatic detection method for small objects in high-resolution images based on computer vision and deep learning. Then, the low-resolution detectors are trained at different scales and applied for detection, and the detection results at different scales are obtained. Finally, these detection results are fused to obtain the final detection results for small targets. The invention solves the problem that it is difficult for the target detector in the prior art to detect tiny targets in high-resolution images.

Description

A self-portrait of small objects in high-resolution images based on computer vision and deep learning Motion detection method

technical field

The invention belongs to target detection technology, in particular to an automatic detection method for tiny targets in high-resolution images based on computer vision and deep learning.

Background technique

Currently widely used target detectors based on deep learning can be mainly divided into two categories: the first category is two-stage target detectors, such as Fast R-CNN, Faster R-CNN, Mask R-CNN, etc. The characteristics of these algorithms are to divide the target detection into two stages: first extract the candidate area, and then send it to the detection network to complete the target location and recognition. The second category is one-stage target detection algorithms, such as Single Shot Detection (SSD), You Look Only Once (YOLO), YOLO 9000, YOLO V3, etc. Such algorithms do not need to pre-extract candidate frames, but directly It is an end-to-end target detection algorithm to complete the target position regression and category judgment through the preset frame in the network. In the scene where the scale of the target to be detected is large and not dense, both the two-step target detector and the one-step target detector have higher detection accuracy, while the latter has a faster detection speed than the former. However, due to the influence of the actual size of the target, shooting equipment, shooting distance, observation scale and other factors, the real target often appears as a small target in the image. Compared with large objects, small objects have fewer pixels and less obvious features to extract. When detecting these small targets, it is difficult to achieve better detection results with either two-step or single-step detectors.

At present, the optimization of the small target detection algorithm mainly focuses on the improvement of the model, that is, on the premise that the input low-resolution image size remains unchanged, the feature extraction ability and detection accuracy of the detector can be improved by improving the model structure. The most effective improved algorithm at present is Feature Pyramid Networks. The network can be embedded in the above two-step, single-step detector, which can fuse the low-level feature map and high-level feature map generated by the main network in a specific way to complete the reconstruction of the feature pyramid. After this operation, the receptive field range of the low-level feature map is improved, and its semantic information is enhanced, which finally greatly improves the accuracy of the model for small target detection.

Although the above improvements can improve the model detection accuracy, the objects processed by these models are still low-resolution images. With the improvement of the hardware performance of camera equipment, people can obtain higher resolution images. Compared with low-resolution images, small objects can be represented by more pixels in high-resolution images, that is, they can be more clearly depicted. This feature provides effective data support for small target detection tasks. Although supported by the data, the current detection algorithms are basically not suitable for images with a resolution of up to tens of millions of pixels. However, if the high-resolution image is downsampled to adapt to the detection model, information will be lost, the characteristics of the high-resolution image cannot be fully utilized, and it is difficult to detect small objects in the end. Aiming at the problem of small target detection in high-resolution images, the target detection process based on high-resolution satellite images Satellite Imagery Multiscale Rapid Detection with Windowed Networks (SIMRDWN) uses a fast detector to detect candidate regions obtained through sliding windows, which can complete any A fast detection task for high-resolution images of size. However, this method has low detection accuracy, many false alarms, and long execution time. In response to this problem, this patent proposes a simple and effective method to solve the problem of small target detection in high-resolution images. The algorithm splits the original detection task on different scales of the image to obtain a multi-scale detection task group with logical correlation; then trains corresponding low-resolution object detectors for detection tasks at different scales; The following detection results are fused to obtain the final defect mark detection result. The purpose of the research is to construct a pattern recognition framework based on deep neural network and explore small target detection methods based on high-resolution images.

SUMMARY OF THE INVENTION

Aiming at the problems in the prior art, the present invention provides an automatic detection method for small objects in high-resolution images, which solves the problem that it is difficult for object detectors in the prior art to detect tiny objects in high-resolution images.

The method is divided into detection and training processes, an automatic detection method for small targets in high-resolution images based on computer vision and deep learning, and is characterized in that the steps are as follows:

S1 detection process

S1.1 Establish a multiscale task force

According to the Gaussian pyramid theory, a dual-scale image pyramid is established for the original high-resolution image to be detected, and the small target detection task is appropriately decomposed at these two scales to obtain a multi-scale task group. Among them, the large-scale object segmentation task that has no inclusion relationship with the small object is set at the large scale, and the original small-object detection task is set at the small scale.

S1.2 Object segmentation at large scale

At large scales, the Mask R-CNN model is used to segment large-scale images, and the resulting low-resolution masks are upsampled to restore the original image resolution.

S1.3 Target detection at small scale

In the small scale, the overlapping sliding window is used to extract the candidate regions in the small scale image, and the candidate regions are screened according to the mask, and the candidate regions that have no intersection with the large target region in the Mask are sent to the target detector SSD for detection. After the detection of all candidate regions is completed, these detection results are mapped from the sub-regions back to the original image.

S1.4 Fusion of segmentation and detection results under dual scales

Using the segmentation mask obtained at the large scale, the detection frame obtained at the small scale is used for secondary screening. First, the segmentation mask is morphologically processed, then the detection frames appearing in the large target area are deleted, and finally, the overlapping detection frames are fused by non-maximum suppression to obtain the final detection results of small targets in high-resolution images.

S2 training process

S2.1 Segmentation model training

Using large-scale images and annotation information, the Mask RCNN model is trained by transfer learning, and the best performing node on the validation set is saved as the trained segmentation model Model _S.

S2.2 Initial training of detection model

The defect surrounding area in the small-scale high-resolution image is randomly cropped in a specific way to obtain a slice sample set that conforms to the input size of the SSD model. The SSD model is trained by transfer learning, and the best performing node on the validation set is saved as the trained segmentation model Model _D1 .

S2.3 Secondary training of detection model

The trained Model _S and Model _D1 are embedded in the above detection process, and the high-resolution atlas is detected according to them. The false detection frame in the result is cut out as a separate class and added to the original slice set, and the detection model Model _D1 is retrained using the new training set to obtain the secondary training model Model _D2 . Finally, the target detector Model _D1 in the original detection framework is replaced with the secondary training model Model _D2 to complete the construction of the final detection framework.

To sum up, the present invention uses a combination of multi-scale segmentation and detection model to complete the automatic detection of small targets in high-resolution images. At the same time, the main work of the present invention is to combine the deep learning model to understand the picture content on multiple scales, and to use different information to improve the detection accuracy and speed up the detection efficiency.

Description of drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without any creative effort.

1 is a schematic flowchart of a method for detecting small objects in a high-resolution image proposed by the present invention;

2 is a flowchart of a method for detecting small defects in a high-resolution image of a floor provided by an embodiment of the present invention;

3 is a high-resolution image of a floor including defects provided by an embodiment of the present invention;

FIG. 4 is a large-scale non-wall area segmentation algorithm provided by an embodiment of the present invention;

FIG. 5 is a schematic diagram of an in-window sampling method provided by an embodiment of the present invention;

FIG. 6 is a defect detection algorithm at a small scale provided by an embodiment of the present invention;

FIG. 7 is a comparison of single-scale and multi-scale fusion detection effects provided by an embodiment of the present invention. (a) is the detection result on the small scale; (b) is the small scale detection result after the fusion segmentation mask;

FIG. 8 is a training result of Mask-RCNN provided by an embodiment of the present invention. (a) is the mAP change on the validation set; (b) is a schematic diagram of the model detection effect;

FIG. 9 is a training result of an SSD provided by an embodiment of the present invention. (a) is the mAP change on the validation set; (b) is a schematic diagram of the model detection effect;

FIG. 10 is the performance of the detection algorithm before and after the secondary training provided by the embodiment of the present invention on the images of the validation set. (a) is the test result 1 before the second training; (b) is the test result 1 after the second training; (c) is the test result 2 before the second training; (d) is the test result after the second training 2.

Detailed ways

In order to better explain the present invention and facilitate understanding, the present invention will be described in detail below with reference to the accompanying drawings and through specific embodiments.

In the following description, various aspects of the present invention will be described, however, to those of ordinary skill in the art, the present invention may be practiced using only some or all of the structures or processes of the present invention. For clarity of explanation, specific numbers, configurations, and orders are set forth, but it will be apparent that the invention may be practiced without these specific details. In other instances, well-known features have not been described in detail in order not to obscure the present invention.

Aiming at the problem that most of the existing small target detection technologies are aimed at model improvement and are difficult to process high-resolution images, the present invention proposes an automatic detection method for small targets in high-resolution images based on computer vision and deep learning. By relying on multi-scale task decomposition and combining large-scale segmentation and small-scale detection results, the automatic detection of small objects in high-precision and high-speed high-resolution images is finally realized.

1. Detection process

(1) Establish a multi-scale task force

According to the Gaussian pyramid theory, an image pyramid with two scales, one large and one small, is established for the original high-resolution image to be detected. Large-scale images are first obtained by Gaussian filtering of the original image, and then 8 times downsampling. The width and height of the generated large-scale image are both 1/8 of the original image.

The two-dimensional Gaussian kernel function is as follows:

Use the Gaussian kernel template G(x, y) of size 5*5 to perform convolution operation with the original image I(x, y), and perform 8 times downsampling to obtain a large-scale image L(x, y), the formula It is expressed as follows:

The small scale is the original scale of the image, so the small scale image S(x, y)=I(x, y) can be obtained without any operation on the original image.

For a large-scale image L(x, y), a target segmentation task is set. The selection of segmentation targets should be large targets that do not contain small targets to be detected, and these targets can still retain the main features in the large-scale image. For the small-scale image S(x, y), the small target detection task is still set.

(2) Target segmentation at large scale

At large scale, the trained Mask R-CNN model is used to segment the large target set in the large scale image, and the binary mask image Ms with the same resolution as the large scale image L(x, y) is obtained. (x, y). In order to restore the resolution size of the original image I(x, y) to facilitate the use of subsequent steps, it is also necessary to upsample the obtained low-resolution mask Ms(x, y) by 8 times to obtain M(x, y) .

M(x,y)=8↑Ms(x,y)

The interpolation method adopts the nearest neighbor interpolation method. In the obtained high-resolution mask image M(x, y), the pixel value of the large target area is 255, and the pixel value of other areas is 0. At the same time, segmentation methods at large scales are not limited to depth models, but can also be supplemented by traditional image processing methods.

(3) Target detection at small scales

At small scales, since the image resolution is much higher than the input size of the SSD detector, a sliding window that is consistent with the input size of the detector is used to overlap and slide on the small-scale image to extract candidate detection regions in the small-scale image. Proposal _S. Before being sent to the SSD for detection, the candidate regions are first screened according to the mask _M (x, y). If the 9 preset sampling points in Proposal _M are all 255 pixels, that is, the area is considered to be mainly a large target area, then subsequent detection will not be performed. After the window traverses the entire image, the detection result corresponding to each Proposal _S will be obtained. There are D ₁ detection results that are greater than the confidence threshold, and the d-th detection box can be expressed as coordinates (x _d , y _d , w _d , h _d ), category _{cd and confidence s d ,} then the detection set DBox ₁ ={((x _d , y _d , w _d , h _d ), c _d , s _d )|d∈[1, D ₁ ]}.

These detection results are then mapped from the Proposal _S back to the original image. The mapping method is as follows: Let the position of the upper left corner of the Proposal _S in the original image be recorded as (X _ps , Y _ps ). Then according to the coordinate transformation, the four parameters (X _d , Y _d , W _d , H _d ) of the target detection frame corresponding to the original image can be obtained

X _d =X _ps +x _d

Y _d =Y _ps +y _d

W _d = w _d

H _d = h _d

The detection set becomes DBox ₂ ={((X _d , Y _d , W _d , H _d ), c _d , s _d )|d∈[1, D ₁ ]}.

(4) Fusion of segmentation and detection results under dual scales

After completing the multi-scale segmentation and detection, the multi-scale information can be fused to improve the detection accuracy of small objects. Using the segmentation mask obtained at the large scale, the detection result set DBox ₁ obtained at the small scale can be used for secondary screening. First, the segmentation mask is etched so that the area of the large target area is reduced. Then it will be judged whether the pixel values of the four corners of the detection frame are all 0, if so, keep it, otherwise the detection frame will be discarded. Assuming that the number of filtered detection results becomes D ₂ , the detection set becomes DBox ₂ ={((X _g , Y _g , W _g , H _g ), c _g , s _g )|g∈[1, D ₂ ]}.

In order to suppress the overlapping phenomenon of the detection frames caused by the overlapping sliding of the sliding window, the non-maximum value suppression algorithm is used to complete the further fusion of the detection frames, and a set containing F final detection results is obtained DBox _final = {((X _f , Y _f , W _f , H _f ), c _f , s _f )|g∈[1, F]}. Now suppose there is a set B of candidate boxes and its corresponding set of scores S, the algorithm of non-maximum suppression is as follows:

1. Find the element dbox with the highest confidence s in DBox ₂ ;

2. Delete dbox from DBox ₂ and add it to DBox _final ;

3. Delete other boxes whose overlapping area with the dbox coordinate frame is greater than the threshold N _t from DBox ₂ ;

4. Repeat steps 1-3 above.

2. Training process

(1) Segmentation model training

The instance segmentation model Mask RCNN is trained with a large-scale image set and its annotation information. First, the sample set is divided into two subsets, training and validation. Then use the training set to adjust the model parameters, and test the pros and cons of the model on the validation set every 10 steps. The training method adopts transfer learning, that is, pre-training on a large data set, and then fine-tuning the parameters on a small training set of practical application scenarios. The optimization method adopts Mini-Batch GradientDescent, and the batch size is set according to the hardware environment. After training for a specified number of rounds, the best performing node on the validation set is finally saved. Get a Model _S.

(2) Initial training of samples of the detection model

Unlike the training of the segmentation model, the training of the detection model requires the samples to be processed first. The defect surrounding area in the small-scale high-resolution image is randomly cropped to obtain a slice sample set that conforms to the input size of the SSD model. In order to ensure that the annotation boxes in the slice are complete and not truncated, a random cropping algorithm is designed as follows. Let W _w and H _w be the length and width of the window, let the length and width of the label box BBox _i be W _bi and H _bi , and the coordinates of the upper left corner point are (x _bi , y _bi ), where i∈[0, N] . The algorithm flow is as follows:

1. i=0

2. For BBox _i , the influence of the nearby annotation frame is not considered first, and only the cropping box window is required to completely contain BBox _i , then the coordinate value range of the upper left corner of the cropping window is a rectangular area. The coordinates of the upper left corner of the area Cand are (x _bi +W _bi -W _w , y _bi +H _bi -H _w ). The length and width of this region are (W _w -W _bi , H _w -H _bi ). Set up a mask image to record the positions of selectable points with points on which the pixel value is 255. At the same time, it is also necessary to calculate that the upper left corner of Cover, which may be covered by a window, is the same as Cand, the width is W _cover =2W _w -W _bi , and the height is H _cover =2H _w -H _bi . Take the BBox _i area as the center, extend its side length, and divide the rest of the Cover area into 8 rectangular blocks. Starting from the upper left block, assign serial numbers to the blocks clockwise, as follows: upper left block - 0, upper right block - 1, upper right block - 2, right right block - 3, lower right block - 4, the lower block - 5, the lower left block - 6, the positive left block - 7.

3. Traverse all the annotation boxes and record the M detection boxes (except BBox _i ) that intersect the Cover area {NearBBox _j |j∈[0, M]}

4. Traverse {NearBBox _j | j∈[0, M]}, and set 0 to different parts of CandMask according to the position of the sub-block where the intersection part is located. Taking the upper left corner of each subregion as (0, 0), the coordinates of the upper left corner of the intersection of NearBBox _j and CoverSubRegion _k are (x _{LT, jk} , y _{LT, jk} ), and the coordinates of the lower right corner are (x _{RD , jk} , y _{RD, jk} ). The specific operation method is as follows:

5. At this time, all positions with a pixel value of 255 in CandMask are candidate points that satisfy the coordinates of the upper left corner of the cropping frame that can completely contain the annotation frame. Then some random coordinates can be picked from it and used to crop the image. At the same time, the coordinate information of the callout frame should be adjusted linearly.

After the sample set for training the detector is produced according to the above algorithm, the sample set is firstly divided into two subsets of training and verification. Then use the training set to adjust the model parameters, and test the pros and cons of the model on the validation set every 10 steps. The training method adopts transfer learning, that is, pre-training on a large data set, and then fine-tuning the parameters on a small training set of practical application scenarios. The optimization method adopts Mini-Batch Gradient Descent, and the batch size is set according to the hardware environment. After training for a specified number of rounds, the best performing node on the validation set is finally saved. Get Model _D1 .

(3) Secondary training of detection model

The above detection process is used to detect the high-resolution atlas, and the false positive frame is added to the original slice set as a separate category, and the detection model is retrained to obtain a secondary training model. Then replace the object detector in the original detection framework with this secondary training model. According to the area intersection of the detection frame and the real frame, it is judged whether a detection result is a false detection. Suppose the detection threshold is T _IOU , then

Taking the error detection frame as the center, it is randomly cropped by the method in two (two) to obtain a negative sample set that matches the input size of the detector. Then remove the original detection frame on each slice, and reset a detection frame whose category is normal and whose size is the size of the slice. Then, these annotated slices are mixed with the initial positive sample slice set, and the initial training model Model _D1 is retrained with the synthesized data set to obtain Model _D2 . Finally, Model _S and Model _D2 are used for detection in the detection process.

In summary, the present invention disassembles the small target detection task in high-resolution images on multiple scales, and uses different depth models to solve each sub-task, and finally fuses the multi-scale segmentation and detection results to obtain high-precision detection results. Test results.

Example

Figure 1 shows a general detection framework for small objects in high-resolution images based on computer vision and deep learning. FIG. 2 shows a schematic diagram of the process of applying the detection framework to the automatic detection of small defects in high-resolution floor images in this embodiment and an intermediate result.

The embodiment adopts the real shot image of the wall of a certain district, the resolution is 7952*5304, and the file size is generally about 30M. Initially obtained 107 defect images and 195 defect targets. As shown in Figure 3, the right frame is a partially enlarged defect.

First, the detection process is introduced. According to the first detection step, a dual-scale pyramid is first constructed, as shown in Figure 2. On a large scale, non-wall regions that do not contain defects are set as segmentation targets, such as windows, air conditioners, and sky. On a small scale, common floor defects such as missing bricks and broken bricks are used as detection targets.

According to the detection step 2, Mask RCNN is used to segment the windows and air conditioners on a large scale, and the sky is segmented by the traditional area growth method, and finally a non-wall large target segmentation mask is obtained, as shown in Figure 4.

According to the detection step 3, the SSD is used to detect the missing bricks on a small scale. The candidate frame proposed by the sliding window of size 640×640 is initially screened by using the wall mask. Among them, the positions of 9 preset sampling points in Proposal _M are shown in Figure 5. Then use SSD to detect the sub-block, and map the coordinates of the detection frame back to the original image. The process is shown in Figure 6. The defect detection results on the small scale are shown in Fig. 7(a).

According to the fourth detection step, the results of large-scale segmentation and small-scale detection are integrated to improve detection accuracy. The final detection result after mask secondary screening and non-maximum suppression processing is shown in Figure 7(b).

Next, the training process is introduced. According to the first training step, 31 samples with a resolution of 994×663 used to train the instance segmentation model MaskR-CNN are divided into training set (25) and validation set (6). The number of windows and air conditioners in this picture varies. Based on the COCO pretrained model parameters, the parameters are fine-tuned on the above training set. The batch size is set to 1, the initial learning rate is set to 0.0001, and the number of iterations is 100000 steps. During the training process, the variation of the mean mean precision mAP of the model on the validation set is shown in (a) of Figure 8. The mAP curve changes on the validation set are shown in Fig. 8(a). The model with mAP of 96.89% in step 38440 is saved as the final segmentation model, and its confidence threshold is set to 0.5, and some of the detection results are obtained as shown in (b) of Figure 8.

According to the second training step, a 640×640 slice atlas is first created from the high-resolution image set. Among them, there are 107 samples used to train the SSD model of the single-step detector, and the image resolution is generally 7952×5304, which are divided into training set (88), validation set (9) and test set (10). The number of defects in the pictures varies, and the number of slice samples corresponding to the generated training set, validation set, and test set are 3180, 220, and 200, respectively. Then, the pre-trained SSD model on the COCO dataset is used to perform migration learning on the sample set of floor defect slices. The batch size is set to 4, the learning rate is set to 0.00005, and the iteration is 30,000 times. The mAP of the model on the validation slice set as a function of the number of iterations is shown in the figure, where the IOU threshold is set to 0.5. It can be found from (a) of Figure 9 that after 100,000 iterations, the mAP of the model on the validation slice set can reach about 0.7, and then tends to be stable. The model at the highest mAP node on the validation slice set is saved as the initial training model, and the mAP of this model on the test slice set is 0.714. The confidence threshold of the initial training model is set to 0.5, and some of the detection results are shown in Figure 9(b).

According to the training step 3, the Mask RCNN and SSD for the initial training are used to detect the high-resolution images of the training and validation sets according to the above detection process, and the obtained negative samples are cropped and added to the original slice set. Perform secondary fine-tuning and replace the original SSD with it to complete the final construction of the detection framework. Figure 10 shows the comparison of the detection results of the first training SSD and the second training SSD in the detection algorithm.

Finally, it should be noted that the various parameters designed by this method need to be adjusted according to the specific interests of the actual application. The above-mentioned embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand: The technical solutions described in the embodiments are modified, or some or all of the technical features thereof are equivalently replaced; and these modifications or replacements do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present invention.

Claims (6)

1. A method for automatically detecting small targets in a high-resolution image based on computer vision and deep learning is characterized by comprising the following steps:

s1 detection flow

S1.1 establishing multiscale task groups

Establishing a dual-scale image pyramid for the original high-resolution image to be detected according to a Gaussian pyramid theory, and decomposing a small target detection task under the two scales to obtain a multi-scale task group; setting a large target segmentation task which has no inclusion relation with a small target at a large scale, and setting an original small target detection task at a small scale;

s1.2 object segmentation at large scale

Under the large scale, a Mask R-CNN model is used for segmenting the large scale image, and the obtained low resolution Mask is subjected to up-sampling to recover the resolution of the original image;

s1.3 Small Scale target detection

Under the small scale, extracting a candidate region in the small scale image by using an overlapping sliding window, screening the candidate region according to a Mask, and sending the candidate region without intersection with a large target region in the Mask to a target detector SSD for detection; after all the candidate areas are detected, mapping the detection results from the sub-areas back to the original image;

s1.4 segmentation under dual scales and detection result fusion

Carrying out secondary screening on the detection frame obtained under the small scale by utilizing the segmentation mask obtained under the large scale; firstly, performing morphological processing on a segmentation mask, then deleting a detection frame appearing in a large target area, and finally fusing overlapped detection frames by applying non-maximum value inhibition to obtain a final detection result of a small target in a high-resolution image;

s2 training procedure

S2.1 segmentation model training

The Mask RCNN Model is trained in a transfer learning mode by utilizing pictures and labeled information under large scale, and nodes stored on a verification set are used as a trained segmentation Model_S；

S2.2 detection model initial training

Randomly cutting a defect peripheral area in the small-scale high-resolution image according to a specific mode to obtain a slice sample set which accords with the input size of the SSD model; training an SSD Model in a migration learning mode, and saving the best-performing node on a verification set as a trained segmentation Model_D1；

S2.3 test model second training

Model to be trained_SAnd a Model_D1Embedding the image into a detection flow, and detecting according to the high-resolution atlas; cutting out false detection frames in the result, adding the false detection frames as a single class into the original slice set, and detecting the Model by using the new training set_D1Retraining to obtain a secondary training Model_D2(ii) a Finally, Model of secondary training Model is used_D2Model for replacing target detector in original detection frame_D1And finishing the construction of the final detection frame.

2. The method for automatically detecting small objects in high-resolution images based on computer vision and deep learning according to claim 1,

at small scale, since the image resolution is much higher than the input size of the SSD detector, use is made ofA sliding window with the same size as the input of the detector is overlapped and slid on the small-scale image to extract the candidate detection region Proposal in the small-scale image_S(ii) a Before the SSD is detected, screening the candidate area according to a mask M (x, y), wherein the screening mode is as follows: simultaneous extraction of the sub-regions Proposal on M (x, y) using one and the same sliding window_MIf Proposal is used_MThe middle 9 preset sampling points are all 255 pixels; after the window traverses the complete picture, each Proposal will be obtained_SCorresponding detection results; setting the detection result greater than the confidence threshold to have D₁The d-th detection frame is expressed as a coordinate (x)_d，y_d，w_d，h_d) Class c_dAnd a confidence s_dThen detect the set DBox₁＝{((x_d，y_d，w_d，h_d)，c_d，s_d)|d∈[1，D₁]}；

These measurements were then derived from Proposal_SMapping back to the original image, the mapping method is as follows: proposal is provided_SThe position of the upper left corner in the original image is recorded as (X)_ps，Y_ps) (ii) a Then 4 parameters (X) of the target detection frame corresponding to the original image can be obtained according to the coordinate transformation_d，Y_d，W_d，H_d)

X_d＝X_ps+x_d

Y_d＝Y_ps+y_d

W_d＝w_d

H_d＝h_d

Detection ensemble becomes DBox₂＝{((X_d，Y_d，W_d，H_d)，c_d，s_d)|d∈[1，D₁]}。

3. The method for automatically detecting small targets in high-resolution images based on computer vision and deep learning as claimed in claim 1, wherein after segmentation and detection under multiple scales are completed, multi-scale information is fusedThe detection precision of the small target is improved; utilizing the segmentation mask obtained under the large scale to detect the DBox result set obtained under the small scale₁Carrying out secondary screening; firstly, carrying out corrosion treatment on the segmentation mask to reduce the area of a large target region; then judging whether the pixel values of 4 angular points of the detection frame are all 0, if so, keeping, otherwise, abandoning the detection frame; assuming that the number of the test results after screening is D₂If yes, the detection set becomes DBox₂＝{((X_g，Y_g，W_g，H_g)，c_g，s_g)|g∈[1，D₂]}。

4. The method as claimed in claim 1, wherein the method for automatically detecting small objects in the high resolution image based on computer vision and deep learning is characterized in that, in order to suppress the overlapping phenomenon of the detection frames caused by the overlapping sliding of the sliding window, a non-maximum suppression algorithm is applied to complete the further fusion of the detection frames, so as to obtain a set DBox containing F final detection results_final＝{((X_f，Y_f，W_f，H_f)，c_f，s_f)|g∈[1，F]}。

5. The method for automatically detecting the small and medium targets in the high-resolution image based on the computer vision and the deep learning as claimed in claim 1, wherein a segmentation model Mask RCNN is trained by a large-scale picture set and labeling information thereof; firstly, dividing the sample set into two subsets of training and verification; then, adjusting model parameters by using the training set, and testing the quality of the model on the verification set every 10 steps; the training mode adopts transfer learning, namely pre-training is carried out on a large data set, and then parameter fine tuning is carried out on a small training set of an actual application scene; training the designated number of rounds by adopting a small batch gradient descent method, and finally saving the nodes on the verification set to obtain a Model_S。

6. Automatic detection of small objects in high-resolution images based on computer vision and deep learning as claimed in claim 1The measuring method is characterized in that the peripheral area of the defect in the small-scale high-resolution image is randomly cut to obtain a slice sample set which conforms to the input size of the SSD model; in order to ensure that all the labeled boxes in the slice are complete and not truncated, a random clipping algorithm is designed as follows; let W_wAnd H_wThe length and width of the window are provided with a marking frame BBox_iHas a length and width of W_biAnd H_biThe coordinate of the upper left corner point is (x)_bi，y_bi) Wherein i ∈ [0, N]。

Images