现代电子技术

为同时捕捉不同尺度的特征，精准区分前景手势和背景干扰，文中提出基于多尺度卷积神经网络的连续手语精准识别方法，旨在解决手势多样性带来的识别难题。利用主导手轨迹信息的手语语句分割算法，检测连续手语视频中的过渡动作，分割连续手语视频，得到多个复合视频段；多尺度卷积神经网络通过大小不同的卷积核，同时捕捉每个复合视频段不同尺度的特征，精准区分前景手势和背景干扰；利用多尺度空洞卷积池化金字塔模块融合各复合视频段的多尺度特征，充分利用手语动作的多尺度信息，增强网络对手势多样性的处理能力；采用Softmax分类器处理融合多尺度特征，得到各复合视频段的手语精准识别结果；按照时间先后顺序串联识别结果，得到最终的识别结果。实验结果证明，所提方法可精准识别连续手语，且在不同背景干扰情况下的连续手语识别的决定系数与1较为接近，即连续手语识别精度较高，可以有效解决连续手语识别中的难点。

To capture features of different scales simultaneously and distinguish foreground gestures from background interference accurately, a continuous sign language accurate recognition method based on multi ⁃ scale convolutional neural networks is studied, aiming to solve the recognition difficulties caused by gesture diversity. A sign language sentence segmentation algorithm utilizing dominant hand trajectory information is used to detect transitional actions in continuous sign language videos, segment continuous sign language videos, and obtain multiple composite video segments. Multi ⁃ scale convolutional neural networks are used to capture features of different scales in each composite video segment by convolution kernels of different sizes, so as to distinguish foreground gestures from background interference accurately. A multi⁃scale dilated convolution pooling pyramid module is used to fuse the multi⁃scale features of each composite video segment, and the multi⁃scale information of sign language actions are fully utilized to enhance the network's ability to handle gesture diversity. A Softmax classifier is used to process and fuse multi⁃scale features, and accurate sign language recognition results for each composite video segment are obtained. The recognition results are concatenated in chronological order to obtain the final recognition results. Experimental results have shown that the method can recognize continuous sign language accurately, and its determination coefficient of continuous sign language recognition under different background interference conditions is close to 1, indicating high accuracy in continuous sign language recognition. To sum up, the proposed method can effectively solve the difficulties in continuous sign language recognition.