Advanced Applications - Detect Small Targets

1. Principle

1.1 Background

In various scenes, there are often situations where the image is large and the target to be detected is very small, such as detecting small spots or small cracks in the image. In this scenario, bad detection results will be obtained.

To enhance the detection of small targets within a larger image, it is necessary to perform an overall analysis of the entire image in order to identify the target region.

1.2 Calculate the area ratio of the whole image to the detected object

In general, the system will convert the user input image into a standard one megapixel size. The system can identify targets with a pixel area of 1000 (about 30*30) well in standard images.

At this point, the ratio of the entire image to the area of the smallest detected object in the graph is :

The area of the whole image/The area of the minimum detection object = 1,000,000 pixels / 1,000 pixels = 1000

In other words, to ensure a good detection effect, the area ratio should be controlled within 1000.

1.3 According to the area ratios to determinate the method

When the area ratio is between 1000 and 5000, please contact the background personnel (Email Address: contact@neurobot.co) to improve the detection effect by adjusting the image resolution. This method is simple and easy to implement, but it cannot solve images with area ratio above 5000.

When the area ratio is above 5000, please use the "crop image" function. The instructions are as follows:

2. Illustration of image segmentation

2.1 Online model development

When creating a new model, select "Crop image". (The image segmentation function only supports target positioning and pixel segmentation, and does not support OCR recognition.)

Determine the cutting method, cutting suggestions:

1. The small picture after segmentation should be as square as possible. That is, if the original picture is square, the two directions are cut to the same number of copies; If the original drawing is thin and long, the number of transverse cutting drawings is slightly more, so that the small drawing is as square as possible
2. After cutting into small maps, ensure the area ratio:
   Area of the whole picture/Area of the minimum detected object < 1000

Click on "Dataset", and then click on "Upload and crop".

Please note the segmentation requirements and upload pictures.

This splits the big picture into smaller ones.

Then the pictures can be marked, trained and tested according to the normal process :-).

2.2 Offline picture test

As the model is developed online, the images are split into smaller images. Therefore, when offline deployment is carried out, please also slice the measured picture according to the online segmentation method to obtain good detection results.