Method and apparatus for conserving computing resources required in the process of verifying the integrity of convolution parameters using a test pattern and improving defect tolerance and swing robustness in extreme situations

摘要

PROBLEM TO BE SOLVED: To save the computing resource required in the process of verifying the integrity of a convolution parameter using a test pattern. A background prediction unit is obtained when information about the (t-2)th image and the (t-1)th image corresponding to the (t-i2)th frame and the (t-i1)th frame of an image is obtained. , The information about the (t−2)th image and the (t−1)th image is referred to, the tth background prediction information of the tth image is generated, and when the tth image is acquired, the pattern insertion unit, With reference to the t-th background prediction information, a test pattern is inserted into the t-th image to generate a verification input, and a neural network is used to apply a convolution operation to the verification input to generate a verification output. The verification unit references the verification output and the reference output corresponding to the test pattern to determine the integrity of the neural network. [Selection diagram] Figure 2