Вплив логарифмічної трансформації вхідних активацій у згорткових мережах на локалізацію ключових точок обличчя

Abstract

The paper considers the application of the logarithmic perception principle in convolutional neural network models, according to which the system&039;s response is determined by relative signal changes. Within this approach, the impact of logarithmic transformation of input activations in the first convolutional layer on facial landmark localization accuracy and model robustness to brightness variations is investigated. Experimental validation on the WFLW dataset using the ResNet-34 architecture demonstrated that such a transformation does not significantly affect localization accuracy under normal brightness conditions; however, it slightly increases the model&039;s robustness to illumination decrease. Specifically, it was found that applying logarithmic transformation reduces the Normalized Mean Error (NME) compared to the baseline model by an average of 0.0019 for a threefold decrease in brightness and by 0.0071 for a sixfold decrease. Thus, the logarithmic transformation of input activations can be considered a tool for enhancing the robustness of convolutional neural networks to input signal intensity variations without increasing their architectural complexity.