Image recognition - Data Acquisition

1 Image acquisition instructions

Calculate the force 、 Models and data are the three elements of AI . When an algorithm model is designed , You need a lot of labeled data to train the machine , Thus making the machine more “ intelligence ”, It can be used in practical application scenarios . If you want the algorithm to further improve performance , You need more refined data to train , Continuous iteration . so to speak ,AI The development of , Data is the foundation , It's also the key .
1.1 image quality
The image target is detected according to the natural features extracted from the target image , Then compare it with the features in the real-time camera image . Although low rated target graphics can usually be well detected and tracked . For best results , Use the following images as much as possible :

surface 1-1 Quality image attribute table

1.2 Camera autofocus

If the target is not well focused in the camera view , Then the camera image result may be blurred and the target details may be difficult to detect . result , Detection and tracking performance may be negatively affected .
It is recommended to use appropriate “ Camera focus mode ” To ensure the best camera focusing conditions .
Please try continuous auto focus mode （FOCUS_MODE_CONTINUOUS_AUTO）, Because it allows your device to automatically adjust the focus as the view changes .

1.3 Dimension of image acquisition

First, consider the actual application scenarios , Consider in the actual application scenario . Such as ： The shooting angle of the assessment scene of the switch machine , The relative position and line of sight angle between the examiner and the switch machine should be within a certain reasonable range （ angle 30°~ 60°, distance 60cm ~ 150cm）. Think about the problem from the perspective of bionics .
Image pixels ： The larger the image pixel , The more picture information saved . But too large image will bring many problems to model training and data , Such as ： Increase the training cost of the model 、 Improve the hardware cost of image acquisition and deployment 、 Reduce the real-time performance of detection and recognition . Temporary use 480p、720P and 1080P Image of .
The following are the dimensions or factors that should be considered when collecting images :

surface 1-2 Image acquisition dimension description table

2 Image acquisition cases

Think about the problem from the perspective of bionics , Collect pictures from a human perspective , Recognize objects with human cognition .

2.1 Image acquisition

What kind of data to collect , It means that our model may learn some characteristics . The data we feed to the model is the scene where we need to detect the target

2.1 The principle of the angle change principle

Think about the problem from the perspective of bionics , Collect pictures from a human perspective . The angle of image acquisition should meet the actual needs , And make adjustments in combination with the corresponding scenes .

General image acquisition requirements are ： Multiple perspectives , Collect pictures in all directions . Pictured 2 Shown , Consider the angle of the camera when collecting images from the perspective of people 、 distance , And the height of the camera itself . Here, the distance from the switch machine is set to （0.8 To 1.5M）. chart 3 The height range of the camera （1.5~1.8M） The tilt angle of the camera is adjusted with distance and height , Ensure that the turning machine is always in the center of the image .

2.2 Turning machine image acquisition

Identify all parts of the turnout turning machine , Component drawings to be identified ：

Try camera height and distance , The camera angle is aligned with the middle position of the turning machine

Data set iteration process table

In the case of small data sets , Want to achieve better recognition effect , According to the usage scenario , Continuously revise the data set . Tagged data , Check it several times before reuse .

Different project requirements , The requirements for samples will also be different , Reasonably according to the needs of the project , Collect appropriate data , Calibrate the appropriate sample . Excessive pursuit of generalization effect , It will cause unnecessary waste .