Detection and recognition of extremely small targets
Object detection is becoming increasingly popular nowadays and there is increasing interest since the use of deep learning methods.
Object detection is becoming increasingly popular nowadays and there is increasing interest since the use of deep learning methods. Nowadays, based on the widespread use of drones and aircraft, applications that use aerial photos are very popular. This approach is more difficult than common object detection tasks due to the much smaller size of the objects, the computational power limitations associated with using edge devices, and the energy consumption due to limited power supplies, as well as the efficiency and inference time associated with real-time applications.
The Association for Unmanned Vehicle Systems International Student Unmanned Aerial Systems Competition (AUVSISUAS) computer vision task, in addition to object (tag) detection, also requires visual analysis of the detected tags to extract tag colors, alphanumeric symbols, and symbols Color and other characteristics.
Some researchers have proven that the newly proposed method is based on the following methods: YOLO algorithm, k-means clustering, and CNN-based alphanumeric symbol classification, and achieved satisfactory results.
The task now is to develop a tag detection system for aerial photos taken from drones, which allows to connect the tag position with GPS and obtain the target type, orientation and color, alphanumeric symbols and their colors.
The newly proposed method uses YOLO algorithm model to detect objects, k-means clustering to group from background, and SqueezeNet to classify alphanumeric symbols. There is no official dataset for the AUVSI SUAS detection task, so a data generator was generated. It is designed to prepare data in the form of natural background. Different shades of grass, sand, concrete, adding different labels in random locations. Generator uses 10 template numbers and 12 colors. 
The picture above is a case of detection. For example, the target on the left is only 14 pixels in size.
new framework approach
Object Detection
The key with aerial photos is the size of the object, which is usually much smaller due to the altitude of the flight. In this method, a lightweight version of YOLOv4 called YOLOv4-tiny-3l is used, which can detect particularly small objects. This solution allows researchers to perform inference at high FPS rates on less powerful devices such as Nvidia Jetson.
Detection metrics with tag size attention
Tag Segmentation
The segmentation task relies on cluster detection of target ROIs using the k-Means algorithm with two cluster centers. This approach allows segmentation of markers from the background. The result of clustering is the label of the two cluster centers and the (R, G, B) color in the form of a background. To classify this value, the nearest template color is calculated using Euclidean normal form. Label colors are assigned to inspect ROI image borders. The results of label segmentation and color classification are shown in the figure below. whaosoft aiot http://143ai.com 
Alphanumeric Sign Classification
The third task required classifying alphanumeric tokens into 36 labels (26 symbols and 10 numbers). For this purpose, the EMNIST dataset is used, which extends the standard MNIST with uppercase letters.
Results of methods used for alphanumeric sign classification
It contains 533,993 training and 89,264 test images. Research shows that the lightweight CNN-SqueezeNet is more accurate and faster than the SiameseNet with Triplet Loss method (as shown in the table above), which also requires a classification algorithm such as KNN.
experiment
AUVSI SUAS Competition extends common inspection tasks to create more complex processes. Therefore, these results need to be encoded in a JSON file as shown below.
{”type ”: ”standard ” ,”latitude”: 52.402477,”longitude ”: 16.953619 ,”orientation”: ”n”,”shape”: ”cross”,”background color ”: ”brown” ,”alphanumeric”: ”V”,”alphanumeric color ”: ”yellow”}
Table 1 above contains the results of the detection task for 10px-50px size images. Table 2 above includes the accuracy scores of the tested classification methods. Receive latitude and longitude data from drone GPS sensors via telemetry. In addition, geometric equations, calibrated technical information and measurements from distance sensors can specify more precise positioning values.