Improving the lightweight FCM-YOLOv8n for steel surface defect detection

In response to the deficiencies of existing steel surface defect detection algorithms in terms of resource consumption, detection accuracy, and efficiency, a lightweight steel defect detection algorithm based on YOLOv8n (FCM-YOLOv8n) is proposed. This algorithm incorporates three principal innovative elements. First, a frequency-aware feature fusion network is utilized to efficiently extract and integrate high-frequency information, reducing computational costs while enhancing detection speed. This network ingeniously integrates an adaptive low-pass filter generator (ALPF), an offset generator, and an adaptive high-pass filter generator (AHPF). The ALPF generator forecasts spatially-variant low-pass filters, which serve to attenuate high-frequency constituents within objects, thereby diminishing intra-class disparities during the up-sampling procedure. The offset generator plays a pivotal role in refining pronounced inconsistent features and tenuous boundaries. It achieves this by substituting inconsistent elements with more congruous ones via resampling. Meanwhile, the AHPF generator functions to augment the high-frequency detailed boundary information that is otherwise lost during down-sampling. Collectively, this fusion paradigm substantially augments feature consistency and sharpens object boundaries. Secondly, a lightweight feature interaction module (Cc-C2f) is restructured to effectively preserve spatial and channel dependencies while reducing feature redundancy, lowering model parameters and computational complexity. The Cc-C2f module integrates the lightweight convolutional additive self-attention mechanism (CDSA) and the lightweight convolutional gated linear unit (CGLU). The CDSA module takes into account both channel and spatial information, and employs fast linear transformation to reduce the number of model parameters and computational complexity. The CGLU module combines local and global information to enhance the network's representational ability. Finally, a multi-spectrum attention mechanism is applied to mitigate feature information loss in the frequency domain, improving the accuracy of detecting complex defects. Experimental results on the Severstal and NEU-DET steel defect datasets show that, compared to YOLOv8n, the FCM-YOLOv8n algorithm achieves a 2.2% and 1.5% improvement in mAP@0.5, respectively, with a 0.5 M and 1.5 G reduction in parameters and computational complexity. The FPS reaches 143 f/s and 154 f/s, respectively, demonstrating excellent real-time performance. The algorithm achieves an optimal balance between detection accuracy, computational cost, and efficiency, providing robust support for edge device applications. Further validation on the GC10-DET dataset shows a 2.9% improvement in mAP@0.5 compared to the baseline model, demonstrating the algorithm's exceptional generalization ability. Through comparative analysis with disparate algorithms, the superiority of the proposed algorithm's performance is further accentuated.