一种改进残差深度网络的多目标分类技术

A Multi-objective Classification Technique Based on Improved Residual Deep Network）

摘要:

由于受场景、视角、光照、尺度变化以及局部变形等因素的影响，对重叠目标、拥挤目标、小目标的识别精度较低，提出了一种改进多支路的残差深度卷积神经网络来提高多目标识别的准确度。首先，在第一个卷积残差块layer1后保留恒等映射的同时，增加一个1×1的短接分支尽可能多的保留原始特征；再平行嵌入一个修改激活函数RELU6的空间_通道注意力机制模块（CBAM）；最后这三个特征图进行融合。融合后的特征层着重关注空间和通道中比较显著的信息，从而增强特征图的特征表达能力，以至于卷积神经网络(CNN)获得更多的判别特征，从而大大提高物体识别精度。在FashionMNIST和Cifar10两个数据集的对比性实验显示改进的resnet50算法是准确性-速度较为折中的目标识别模型。

Abstract:

Due to the influence of scene, visual angle, illumination, scale change and local deformation, the recogni- tion accuracy of overlapping target, crowded target and small target is low. An improved multibranch Resnet50 convolutional neural network is proposed to improve the accuracy of multi-objective recognition. First, while retaining constant maps after the first convolutional residual block layer1; Secondly, add as many 1 × 1 short branch as retaining original features; embed a Space_Channel Attention Mechanism Module (CBAM) that modify the activation function RELU6 in parallel; Lastly, the last three feature graphs are fused. The fused feature layer focuses on the more significant information in space and channels, thus enhancing the feature expression ability of feature graphs, so that Convolutional Neural Network (CNN) can obtain more discriminant features, thus greatly improving the accuracy of object recognition. Comparative experiments which are carried out on fashionmnist and cifar10 data sets shows that the improved resnet50 algorithm is a target recognition model with a compromise between accuracy and speed .