With the widespread adoption of deep neural networks (DNNs) in computer vision, post-training quantization has become a key technique for efficient deployment on resource-constrained devices. However, existing mixed-precision post-training quantization methods suffer from two major limitations: the long time required to compute optimal bit-width configurations for low-bit quantization, and the degradation in test accuracy caused by limited calibration samples. To address these issues, this study proposes a Fisher information-guided mixed-precision post-training quantization algorithm. Starting from the Kullback-Leibler divergence, the quantization sensitivity of each layer is efficiently estimated using the average trace of the Fisher information matrix, and a perturbation cost matrix that integrates quantization error and sensitivity is then constructed. The optimal bit-width allocation problem is then solved via integer linear programming. Furthermore, during post-training, a block-wise reconstruction method with feature distribution generalization is proposed, which uses the statistical information from batch normalization layers in the full-precision model to reconstruct input features, thus increasing the diversity of feature distributions during reconstruction. In addition, a random uniform feature-mixing mechanism is incorporated to diversify quantization errors, alleviating the insufficient generalization capability under limited calibration data. Experimental results on multiple mainstream convolutional neural network architectures on the ImageNet dataset demonstrate that the proposed method can complete mixed-precision bit-width search within a relatively short time and effectively improve the accuracy of low-bit quantized models.