Convolutional neural networks (CNNs), such as the EfficientNet series, balance efficiency and performance in image classification. However, existing models struggle with complex textures, fine-grained discrimination, and class-imbalanced data due to rigid architectures. We propose EfficientNet-BSFT-S, an improved model with two key modules. The Block module enhances multi-scale feature extraction by splitting channels and applying instance normalization. The SFTLayer dynamically adjusts features through conditional convolutions to handle distribution shifts. We set extensive experiments on CIFAR-100, DTD, and IP102 by using ImageNet pretraining. The empirical results show Top-1 accuracies of 86.00% (CIFAR-100), 71.22% (DTD), and 71.11% (IP102), surpassing the baseline EfficientNetV2-S by 0.45%, 1.43%, and 1.07%, respectively. It validates the model’s robustness in fine-grained tasks, texture recognition, and imbalanced scenarios.