CSP-CBAMNet: A Compact Multi-Scale Attention Network for Robust Image Recognition in Complex Visual Environments

Realizing image recognition in complex and changeable real scenes is the core problem in the field of computer vision and pattern recognition. The model must meet multiple challenges such as light change, background interference and object appearance diversity. The current mainstream deep learning methods often have performance degradation under such complex conditions, and are often limited by high computing costs, which is difficult to deploy in practice. In response to these problems, this study proposes a compact multi-scale attention network: CSP-CBAMNet. The architecture is specially designed to achieve robust visual recognition under harsh imaging conditions. The architecture adopts Cross Stage Partial (CSP) as the main structure to improve gradient propagation and reduce redundancy. In order to enhance the feature expression ability while maintaining the lightness of the model, the network integrates three key components: the first aspect is Lite-FasterBlock, which realizes efficient feature extraction with lower parameters; the second aspect is the Spatial Pyramid Pooling (SPPF) rapid module, which aims to capture multi-scale contextual information to enrich spatial characterization; and convolutional block attention module (CBAM), which is embedded in the high-level semantic stage to suppress background noise and adaptively focus on the key discrimination area. The comprehensive evaluation on the image classification benchmark shows that CSP-CBAMNet has achieved an accuracy rate of 98.57% under standard conditions. It is worth noting that its robustness is outstanding. Under simulated low-light and noise imaging conditions, the accuracy of the model is maintained at 90.21% and 94.45% respectively. Through systematic ablation experiments, the validity of each proposed module has been further verified. It has been verified that CSP-CBAMNet is a promising deep learning solution, because of its lightweight architecture and strong tolerance to environmental interference, which is especially suitable for deployment on the actual visual recognition platform with limited resources.