With the comprehensive development of intelligent manufacturing, the requirements for quality inspection of industrial products are becoming increasingly stringent. Defect detection on the surface of product images is a key technology in industrial product quality inspection, and its accuracy directly determines the qualification of product quality. When dealing with low-quality or complex images, defect detection accuracy is often poor. To solve this problem, a metal surface defect detection method based on a Mutual Information Adaptive Integrated Neural Network (MI-AINN) is proposed for defect detection. MI-AINN focuses on analyzing and leveraging the features and their spatial distribution across different image channels through a mutual information mechanism. It continuously optimizes the network model through an adaptive sample weight adjustment mechanism, then integrates the sub-models obtained during the optimization process. This approach helps the network to extract deeply complex features, improving detection accuracy. Experimental results show that compared to individual sub-models, MI-AINN achieves the highest F1-score while maintaining both high recall and low false positive rates.