AI noise reduction method for improving night signal-to-noise ratio

Abstract

The invention relates to an AI noise reduction method for improving a night signal-to-noise ratio. The method comprises the following steps: step 1, obtaining original raw data of a night image and carrying out noise reduction processing; 2, performing statistics on gain information of the original raw data, and generating a fusion proportion aiming at the current gain information; 3, performing statistics on motion information of the original raw data, and generating a motion curve fusion proportion aiming at the current motion information; 4, counting the brightness distribution information of the current raw data, and generating a brightness curve fusion proportion for the brightness distribution information; and step 5, generating final raw data after noise reduction based on the data after noise reduction obtained in the step 1 and the three fusion curves obtained in the steps 2-4. The technical problems that the noise form is not good under low illumination, and motion smear is caused in order to solve the noise problem are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent monitoring noise reduction, and relates to an AI noise reduction method, in particular to an AI noise reduction method for improving the signal-to-noise ratio at night. Background Art

[0002] The existing super starlight equipment has poor image quality at night, especially in an environment with an illumination of less than 0.002lux. Moving objects have problems such as high noise, unclear vision, and obvious tailing, and can hardly play an effective monitoring role. After the existing DncNN performs noise reduction, it has a great impact on the clarity in certain scenes at night.

[0003] After searching, no public documents of the prior art that are identical or similar to the present invention were found. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and propose an AI noise reduction method for improving the signal-to-noise ratio at night, which can solve the technical problem that the noise morphology is not good under low light and motion smear will be caused in order to solve the noise problem.

[0005] The present invention solves the practical problem by adopting the following technical solutions:

[0006] An AI noise reduction method for improving the signal-to-noise ratio at night, comprising the following steps:

[0007] Step 1, obtaining the original raw data of the night image for noise reduction;

[0008] Step 2: Count the gain information of the original raw data and generate a fusion ratio for the current gain information;

[0009] Step 3: Count the motion information of the original raw data and generate a motion curve fusion ratio for the current motion information;

[0010] Step 4: Count the brightness distribution information of the current raw data and generate a brightness curve fusion ratio for the brightness distribution information;

[0011] Step 5: Based on the denoised data obtained in step 1 and the three fusion curves obtained in steps 2-4, the final denoised raw data is generated.

[0012] Moreover, the specific method of step 1 is:

[0013] The original raw data of the night image is obtained for denoising, and the raw data is subjected to DncNN denoising and traditional Denoise denoising respectively to generate noise-free data and denoised data;

[0014] Moreover, the calculation formula of the fusion ratio Ka for the current gain in step 2 is:

[0015] Ka=(Bmax-Bmin) / (Gmax-Gmin)*(Gnow-Gmin)+Bmin;

[0016] in

[0017] Bmax is the preset maximum fusion value;

[0018] Bmin is the preset minimum fusion value;

[0019] Gmax is the preset upper limit of the fusion gain (if it is higher than Gmax, use Bmax);

[0020] Gmin is the preset lower limit of fusion gain (no fusion if it is higher than Gmin);

[0021] Gnow is the current gain value;

[0022] Moreover, the specific method of step 3 is:

[0023] Count the motion information of the current raw data, divide the motion information into 32 segments according to size and speed, use the default motion fusion ratio, and generate the motion curve fusion ratio Kb for the motion areas of different segments.

[0024] Kb=(Bi+1-Bi) / (M i+1 -M i )+Bi

[0025] Where Mi+1 is the upper index of the current motion information array;

[0026] Where Mi is the next index of the current motion information array;

[0027] Where Bi+1 is the fusion ratio corresponding to the upper index of the current motion information array;

[0028] Where Bi is the fusion ratio corresponding to the lower index of the current motion information array;

[0029] Moreover, the specific method of step 4 is:

[0030] The brightness distribution information of the current raw data is counted and divided into 32 segments (the brightness is 1024, and each brightness segment interval is 32 brightness levels). The default brightness fusion ratio is used to adjust the fusion ratio for different segments to generate the brightness curve fusion ratio Kc.

[0031] Kc=(Bi+1-Bi) / (L i+1 -L i )+Bi

[0032] Where Li+1 is the upper index of the current brightness array;

[0033] Where Li is the next index of the current brightness array;

[0034] Where Bi+1 is the fusion ratio corresponding to the upper index of the current brightness array;

[0035] Where Bi is the fusion ratio corresponding to the lower index of the current brightness array;

[0036] Moreover, the specific method of step 5 is:

[0037] The noise-free data generated by DncNN in step 1 and the data processed by traditional ISP are combined with the three fusion curves in steps 2-4 to calculate the final fusion ratio α=Ka*Kb*Kc, and then fused according to the alpha fusion formula raw=raw(dncnn)*α+raw(traditional)*(1-α) to generate the final denoised raw data and send it to the subsequent isp module for processing.

[0038] Advantages and beneficial effects of the present invention:

[0039] 1. The present invention proposes an AI noise reduction method for improving the signal-to-noise ratio at night. The AI ​​noise reduction method of DncNN is used to reduce the noise of the camera in a dark environment and improve the imaging effect, thereby reducing the smear problem caused by suppressing noise and improving the clarity of moving objects. A better monitoring effect is achieved.

[0040] 2. The present invention is based on the principle of monitoring effectiveness. The camera automatically identifies the current ambient illumination, distinguishes the gain and degree of motion according to the ambient illumination, and then takes different noise reduction effects into effect according to different brightness ranges to achieve better video monitoring effects.

[0041] 3. The present invention adopts the technical method of integrating DncNN and traditional ISP, which can simultaneously optimize the problem of DncNN's loss of details and the problems of motion smear and reduced clarity that traditional ISP has been unable to solve. The two are integrated and the advantages of the two algorithms are combined to obtain better image effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is a process flow chart of the present invention;

[0043] FIG2( a ) is a comparison diagram of the noise reduction effects of the present invention, traditional ISP, and DncNN - DncNN;

[0044] FIG2( b ) is a comparison diagram of the noise reduction effects of the present invention, traditional ISP, and DncNN - traditional ISP;

[0045] Figure 2(c) is a comparison of the noise reduction effects of the present invention and traditional ISP and DncNN - the present invention.

[0046] FIG3(a) is a local enlarged comparison diagram of the noise reduction effect of the present invention, traditional ISP, and DncNN - DncNN;

[0047] FIG3( b ) is a locally enlarged comparison diagram of the noise reduction effects of the present invention, traditional ISP, and DncNN - traditional ISP;

[0048] FIG3( c ) is a locally enlarged comparison diagram of the noise reduction effects of the present invention and traditional ISP and DncNN - the present invention. DETAILED DESCRIPTION

[0049] The embodiments of the present invention are further described in detail below with reference to the accompanying drawings:

[0050] An AI noise reduction method to improve the signal-to-noise ratio at night, such as Figure 1 As shown, the following steps are included:

[0051] Step 1, obtaining the original raw data of the night image for noise reduction;

[0052] The specific method of step 1 is:

[0053] The original raw data of the night image is obtained for denoising, and the raw data is subjected to DncNN denoising and traditional Denoise denoising respectively to generate noise-free data and denoised data;

[0054] Step 2: Count the gain information of the original raw data and generate a fusion ratio for the current gain information;

[0055] The calculation formula of the fusion ratio Ka of the current gain in step 2 is:

[0056] Ka=(Bmax-Bmin) / (Gmax-Gmin)*(Gnow-Gmin)+Bmin;

[0057] in

[0058] Bmax is the preset maximum fusion value;

[0059] Bmin is the preset minimum fusion value;

[0060] Gmax is the preset upper limit of the fusion gain (if it is higher than Gmax, use Bmax);

[0061] Gmin is the preset lower limit of fusion gain (no fusion if it is higher than Gmin);

[0062] Gnow is the current gain value;

[0063] Step 3: Count the motion information of the original raw data and generate a motion curve fusion ratio for the current motion information;

[0064] The specific method of step 3 is:

[0065] Count the motion information of the current raw data, divide the motion information into 32 segments according to size and speed, use the default motion fusion ratio, and generate the motion curve fusion ratio Kb for the motion areas of different segments.

[0066] Kb=(Bi+1-Bi) / (M i+1 -M i )+Bi

[0067] Mi+1 is the upper index of the current motion information array.

[0068] Where Mi is the next index of the current motion information array

[0069] Where Bi+1 is the fusion ratio corresponding to the upper index of the current motion information array

[0070] Where Bi is the fusion ratio corresponding to the next index of the current motion information array

[0071] The default motion fusion ratio of step 3 is (from motion to stillness):

[0072] [90,90,85,85,80,80,80,70,70,70,60,60,50,50,50,40,40,30,30,30,20,20,20,20,15,15,15,15,15,10,10,10].

[0073] Step 4: Count the brightness distribution information of the current raw data and generate a brightness curve fusion ratio for the brightness distribution information;

[0074] The specific method of step 4 is:

[0075] The brightness distribution information of the current raw data is counted and divided into 32 segments (the brightness is 1024, and each brightness segment interval is 32 brightness levels). The default brightness fusion ratio is used to adjust the fusion ratio for different segments to generate the brightness curve fusion ratio Kc.

[0076] Kc=(Bi+1-Bi) / (L i+1 -L i )+Bi

[0077] Where Li+1 is the upper index of the current brightness array

[0078] Where Li is the next index of the current brightness array

[0079] Where Bi+1 is the fusion ratio corresponding to the upper index of the current brightness array

[0080] Where Bi is the fusion ratio corresponding to the next index of the current brightness array

[0081] The default brightness fusion ratio of step 4 is (from dark to bright):

[0082] [90,90,85,85,80,80,80,70,70,70,60,60,50,50,50,50,50,50,50,50,50,50,50,50,50,50,50,50,50,50,50,50,50,50,50,50,50].

[0083] Step 5: Based on the denoised data obtained in step 1 and the three fusion curves obtained in steps 2-4, the final denoised raw data is generated.

[0084] The specific method of step 5 is:

[0085] The noise-free data generated by DncNN in step 1 and the data processed by traditional ISP are combined with the three fusion curves in steps 2-4 to calculate the final fusion ratio α=Ka*Kb*Kc, and then fused according to the alpha fusion formula raw=raw(dncnn)*α+raw(traditional)*(1-α) to generate the final denoised raw data and send it to the subsequent isp module for processing.

[0086] Figure 2(a) shows the noise reduction effect of only dncnn; Figure 2(b) shows the noise reduction effect of only traditional ISP; Figure 2(c) shows the noise reduction effect of combining dncnn with traditional ISP;

[0087] Figure 3(a)-(c) are partial enlarged views of Figure 2(a)-(c). It can be seen that DNCNN can achieve a good noise reduction effect, with almost no noise and intact dynamic objects. However, there is a relatively large loss of details, especially the details in the bright area; in order to solve the noise problem, the traditional noise reduction is very blurry as a whole, and the moving objects are also blurry and incomplete. The present invention combines the advantages of traditional noise reduction and DNCNN, reduces the noise reduction intensity of traditional noise reduction, and retains more details; combined with the advantages of DNCNN, the integrity of moving objects is retained.

[0088] It should be emphasized that the embodiments of the present invention are illustrative rather than restrictive. Therefore, the present invention includes but is not limited to the embodiments described in the specific implementation modes. Any other implementation modes derived by those skilled in the art based on the technical solutions of the present invention also fall within the scope of protection of the present invention.

Claims

1. An AI noise reduction method for improving the signal-to-noise ratio at night, characterized by: The following steps are involved: Step 1, obtaining the original raw data of the night image for noise reduction; Step 2: Count the gain information of the original raw data and generate a fusion ratio for the current gain information; Step 3: Count the motion information of the original raw data and generate a motion curve fusion ratio for the current motion information; Step 4: Count the brightness distribution information of the current raw data and generate a brightness curve fusion ratio for the brightness distribution information; Step 5: Based on the denoised data obtained in step 1 and the three fusion curves obtained in steps 2-4, the final denoised raw data is generated.

2. The AI ​​noise reduction method for improving the signal-to-noise ratio at night according to claim 1, characterized in that: The specific method of step 1 is: The original raw data of the night image is obtained for denoising, and the raw data is subjected to DncNN denoising and traditional Denoise denoising respectively to generate noise-free data and denoised data.

3. The AI ​​noise reduction method for improving the signal-to-noise ratio at night according to claim 1, characterized in that: The calculation formula of the fusion ratio Ka of the current gain in step 2 is: Ka=(Bmax-Bmin) / (Gmax-Gmin)*(Gnow-Gmin)+Bmin; in Bmax is the preset maximum fusion value; Bmin is the preset minimum fusion value; Gmax is the preset upper limit of the fusion gain (if it is higher than Gmax, use Bmax); Gmin is the preset lower limit of fusion gain (no fusion if it is higher than Gmin); Gnow is the current gain value.

4. The AI ​​noise reduction method for improving the signal-to-noise ratio at night according to claim 1, characterized in that: The specific method of step 3 is: Count the motion information of the current raw data, divide the motion information into 32 segments according to size and speed, use the default motion fusion ratio, and generate the motion curve fusion ratio Kb for the motion areas of different segments. Kb=(Bi+1-Bi) / (M i+1 -M i )+Bi Where Mi+1 is the upper index of the current motion information array; Where Mi is the next index of the current motion information array; Where Bi+1 is the fusion ratio corresponding to the upper index of the current motion information array; Where Bi is the fusion ratio corresponding to the next index of the current motion information array.

5. The AI ​​noise reduction method for improving the signal-to-noise ratio at night according to claim 1, characterized in that: The specific method of step 4 is: The brightness distribution information of the current raw data is counted and divided into 32 segments (the brightness is 1024, and each brightness segment interval is 32 brightness levels). The default brightness fusion ratio is used to adjust the fusion ratio for different segments to generate the brightness curve fusion ratio Kc. Kc=(Bi+1-Bi) / (L i+1 -L i )+Bi Where Li+1 is the upper index of the current brightness array; Where Li is the next index of the current brightness array; Where Bi+1 is the fusion ratio corresponding to the upper index of the current brightness array; Where Bi is the fusion ratio corresponding to the lower index of the current brightness array.

6. The AI ​​noise reduction method for improving the signal-to-noise ratio at night according to claim 1, characterized in that: The specific method of step 5 is: The noise-free data generated by DncNN in step 1 and the data processed by traditional ISP are combined with the three fusion curves in steps 2-4 to calculate the final fusion ratio α=Ka*Kb*Kc, and then fused according to the alpha fusion formula raw=raw(dncnn)*α+raw(traditional)*(1-α) to generate the final denoised raw data and send it to the subsequent isp module for processing.

Images