GradQuant: Low-Loss Quantization for Remote-Sensing Object Detection

Convolutional neural network-based methods have shown remarkable performance in remote-sensing object detection. However, their deployment on resource-limited embedded devices is hindered by their high computational complexity. Neural network quantization methods have been proven effective in compressing and accelerating CNN models by clipping outlier activations and utilizing low-precision values to represent weights and clipped activations. Nonetheless, the clipping of outlier activations leads to distortion of object local features. Furthermore, the lack of enhanced overall feature mining exacerbates the degradation of detection accuracy. To address the limitations above, we propose an innovative clipping-free quantization method called GradQuant, which mitigates the model's quantization accuracy loss caused by clipping outlier activations and the lack of overall feature mining. Specifically, a bounded activation function (sigmoid-weighted tanh, SiTanh) is carefully designed to ensure that object features are represented within a limited range without clipping. On the basis of this, an activation substitution training (AST) method is codesigned to prompt models to focus more on nonoutlier object features instead of outlier-like local ones. Extensive experiments on public remote-sensing datasets demonstrate the effectiveness of the GradQuant method compared with other state-of-the-art quantization methods.