图像显著性检测是计算机视觉的关键分支,它试图模仿人类视觉选择性注意的特性,在图像中精确识别并突出最受关注的目标或区域,影响环境中亮点检测精度的因素包括光照变化、复杂背景、不同的目标尺寸以及低对比度,近年来,以卷积神经网络(CNN)为典型的深度学习技术,虽然大幅提高了检测性能,但在面对复杂背景干扰和低对比度目标时,单一模型的泛化性能仍受到限制。鉴于这些问题,一种依靠多种特征组合来检测图像较大性的新方法,该方法充分融合了多尺度特征提取、跨模态信息融合以及多层次注意力机制,有效提高了模型在复杂背景情况和低对比度环境下的鲁棒性与精确性,实验结果显示,所提出的多特征融合方法在S、F和MAE等主要指标方面的性能有明显提升,准确性和稳定性也有所提高。实验结果基于大量公共标准数据集,并与主流模型的性能进行了对比,本研究还探讨了不同特征与融合策略所起的作用,为复杂场景中的较大性检测研究给予了新的思考方向。Image saliency detection is a key branch of computer vision. It attempts to imitate the characteristics of human visual selective attention, accurately identify and highlight the most concerned target or area in the image. The factors affecting the accuracy of highlight detection in the environment include illumination change, complex background, different target sizes and low contrast. In recent years, convolutional neural network (CNN) as a typical deep learning technology has greatly improved the detection performance, but the generalization performance of a single model is still limited in the face of complex background interference and low contrast targets. In view of these problems, a new method based on a variety of feature combinations to detect the large image is proposed. This method fully integrates multi-scale feature extraction, cross modal information fusion and multi-level atte
