基于深度卷积神经网络的物体识别算法

Object recognition algorithm based on deep convolution neural networks

针对传统物体识别算法中人工设计出来的特征易受物体形态多样性、光照和背景的影响,提出了一种基于深度卷神经网络的物体识别算法。该算法基于NYU Depth V2场景数据库,首先将单通道深度信息转换为三通道;再用训练集中的彩色图片和转换后的三通道深度图片分别微调两个深度卷积神经网络模型;然后用训练好的模型对重采样训练集中的彩色和深度图片提取模型第一个全连接层的特征,并将两种模态的特征串联起来,训练线性支持向量机(Lin SVM);最后将所提算法应用到场景理解任务中的超像素特征提取。所提方法在测试集上的物体分类准确度可达到91.4%,比SAE-RNN方法提高4.1个百分点。实验结果表明所提方法可提取彩色和深度图片高层特征,有效提高物体分类准确度。

Focused on the problem of traditional object recognition algorithm that the artificially designed features were more susceptible to diversity of object shapes, illumination and background, a deep convolutional neural network algorithm was proposed for object recognition. Firstly, this algorithm was trained with NYU Depth V2 dataset, and single depth information was transformed into three channels. Then color images and transformed depth images in the training set were used to fine-tune two deep convolutional neural networks, respectively. Next, color and depth image features were extracted from the first fully connected layers of the two trained models, and the two features from the resampling training set were combined to train a Linear Support Vector Machine （LinSVM） classifier. Finally, the proposed object recognition algorithm was used to extract super-pixel features in scene understanding task. The proposed method can achieve a classification accuracy of 91.4% on the test set which is 4.1 percentage points higher than SAE-RNN （Sparse Auto-Encoder with the Recursive Neural Networks）. The experimental results show that the proposed method is effective in extracting color and depth image features, and can effectively improve classification accuracy.