Image processing method, device and apparatus

An image processing method, comprising:

• calculating a first quantization threshold of activation output layer in a network model of a deep neural network using a saturation mapping method according to a calibrated data set composed of pre-selected images;
• calculating a second quantization threshold of the activation output layer using a non-saturated mapping method according to the calibrated data set;
• performing weighted calculation according to the first quantization threshold value and the second quantization threshold value to obtain an optimal quantization threshold value;
• quantizing the network model according to the optimal quantization threshold value;
• and deploying the quantized neural network to a hardware side and processing the images to be processed, wherein, the performing weighted calculation according to the first quantization threshold and the second quantization threshold to obtain an optimal quantization threshold is specifically:
• S
• sat
• =
• log
• 2
• ⁢
• T
• sat
• 2
• n
• -
• 1
• -
• 1
• ;
• ⁢
• S
• no
• -
• sat
• =
• log
• 2
• ⁢
• 1
• γ
• *
• T
• no
• -
• sat
• 2
• n
• -
• 1
• -
• 1
• ;
• ⁢
• S
• best
• =
• α
• *
• S
• sat
• +
• β
• *
• S
• no
• -
• sat
• ;
• ⁢
• T
• best
• =
• (
• 2
• n
• -
• 1
• -
• 1
• )
• *
• 2
• ceil
• (
• S
• best
• )
• ;
• Tsat is the first quantization threshold value, Tno-sat is the second quantization threshold value, γ is a preset value and is greater than or equal to 1, Ssat is the form of the first quantization threshold value with a base of 2, Sno-sat is the form of the second quantization threshold value with a base of 2, Sbest is a scale factor, Tbest is the optimal quantization threshold value, a is the weight of the first quantization threshold value, p is the weight of the second quantization threshold value, ceil is an upward rounding function, and n is the number of positioning bits needing to be quantized.