Deep learning methods have gained increasing attention in froth flotation monitoring. However, such methods face challenges in perceiving key feature areas and obtaining a sufficient amount of high-quality labeled froth images. To address these issues, a local feature enhancement-based improved semi-supervised condition recognition method (LFE-ISSM) is proposed. First, a local feature enhancement module is constructed by cascading ResNet18 with a convolutional block attention module (CBAM). This dual-attention mechanism, incorporating both channel attention and spatial attention, enhances the expression ability of local features in froth images. Then, an improved semi-supervised algorithm is used to train the condition recognition model, which integrates a threshold-based pseudo-label regularization strategy and a feature similarity alignment strategy within the Mean-Teacher framework to ensure the reliability of pseudo-labels and effectively guide the model to learn the latent structural information in the data. Experimental results show that the proposed method can effectively extract features from froth images and outperforms comparative methods in classification performance on practical industrial zinc flotation datasets.