基于神经网络的多类别目标识别

Multi-category target recognition based on neural network

随着智能化时代的到来,深度学习在图像处理领域的应用越来越广泛,为复杂的图像处理带来了新的解决方法.但是,深度学习在带来高准确推理结果的同时,往往会造成运算量、推理时间以及处理器内存的增加,受限于数据的缺失以及对高性能处理器的依赖.目前市场上有关深度学习技术的相关产品还未被广泛的应用.基于上述问题,提出一种具有广泛应用前景的基于深度学习的多类别目标识别方案,并在国产龙芯派平台下进行测试验证.首先,获取待处理图像数据集,进而在计算机平台下搭建并训练神经网络模型,利用得到的训练参数在龙芯派平台下建立优化后的神经网络结构,并对目标图像进行处理及识别,最后在用户界面显示目标图像所属类别.本系统利用深度学习能够自动从大数据中学习特征的优势,在龙芯派平台下实现上千类生活中常见对象的自动分类,识别准确率高达96%以上,识别速度在3 s以内,且该方案具有优秀的可扩展性.

With the coming of intelligent era,deep learning algorithm has been applied more and more widely in the field of image processing,which brings new solutions to complex image processing.However,while deep learning can produce highly accurate inference results,it often leads to computational complexity,the increase of inference time and processor memory.Due to the lack of data and reliance on high-performance processors,deep learning related products have not been widely used in real life scenarios.Based on the above questions,this paper proposes a multi-category image recognition scheme based on deep learning with broad application prospects and carries out test verification under the Loongson platform.Firstly,the image data set of the category to be identified is obtained,and then the neural network model is built and trained under the computer platform.The optimized neural network structure is established under the Loongson platform using the obtained training parameters,and then the target image is processed and recognized.Finally,the category of the target image is displayed in the user interface.The system utilizes the advantages of deep learning to automatically learn big data features and realizes automatic classification of thousands of common objects in life under the Loongson platform.The recognition accuracy can reach more than 96%,and the recognition time is limited to 3 s or less.The scheme also demonstrates excellent scalability.