This period of time, because contact the related work of the infrared camera image debugging of algorithm and the relationship between image quality and constraint each other, there are many confused and puzzled,

Because do not know much about algorithm, so the algorithm of image request is a little understanding of the word of mouth, such as the sharpness of...

To debug image quality, the relationship between the algorithm and also did the following thinking:
If the algorithm is based on other projects of different sensor trainning before, whether it is necessary for the new sensor from image to trainning? Is deep learning and traditional algorithm two kinds of situations?

Here is an idea, it is assumed that the image quality is a set of optimal parameters, the optimal parameters can satisfy most people look happy, but this group of parameters need to debug out a set of parameters to approach, namely we don't know in advance the optimal parameters of the group, at that time, different people different or the same image quality optimization results, the corresponding parameters, should be around the best parameter distribution, normal distribution can be used to model, optimal parameters should be recognized the group with the highest probability,

At that time, if the training model based on other sensor, an ideal result, the model can and match the sensor image quality fall,

Often there is a strong possibility, however, the corresponding sensor training algorithm, the image quality of matching model often have deviation, at that time, can lead to a consequence, that is, the better the image quality, the algorithm identification effect is poorer,

Then leads to a question: when should continue to optimize the image effect? When should lock the image effect, need to fit the best image effect?

There is need to build a consensus: debug machine algorithm to identify good effect, at least, and the effect of human recognition are relevant, rather than human recognition, the better, the machine identification, back in the opposite direction and line, I think, should eventually recognition of images, the eye looks also should be satisfied,

So, there is a problem here, each image quality first highest visible to human eyes, and then let the machine to fit, this is feasible?

In theory, the algorithm of image effect and human eyes requirements should accord with normal distribution model at the same time,