Visual Confusion Label Tree for Image Classification

Abstract

Convolution neural network models are widely used in image classification tasks. However, the running time of such models is so long that it is not the conforming to the strict real-time requirement of mobile devices. In order to optimize models and meet the requirement mentioned above, we propose a method that replaces the fully-connected layers of convolution neural network models with a tree classifier. Specifically, we construct a Visual Confusion Label Tree based on the output of the convolution neural network models, and use a multi-kernel SVM plus classifier with hierarchical constraints to train the tree classifier. Focusing on those confusion subsets instead of the entire set of categories makes the tree classifier more discriminative and the replacement of the fully-connected layers reduces the original running time. Experiments show that our tree classifier obtains a significant improvement over the state-of-the-art tree classifier by 4.3% and 2.4% in terms of top-1 accuracy on CIFAR-100 and ImageNet datasets respectively. Additionally, our method achieves 124× and 115× speedup ratio compared with fullyconnected layers on AlexNet and VGG16 without accuracy decline.

Enhancing Image Clarity with the Combined Use of REDNet and Attention Channel Module

“…DDT [25] Despite efforts to reduce computational costs, the model may still require significant resources for training and inference. • Output Layer: The output layer merges the input image and the deconvolved image using an addition function.…”

“…DDT [25] is an advanced neural network specifically developed for the purpose of image denoising. It utilises a distinctive architecture that effectively combines local and global data in a simultaneous manner.…”

MMF: Multi-Task Multi-Structure Fusion for Hierarchical Image Classification

“…To obtain a suitable number of superclasses, we perform a series of ablation experiments on our MMF model with a singe label structure H A . Referring to the number of superclasses in H S , We vary the number of H A 's superclasses in [18,20,25,30] for CIFAR100 , and [15,18,20,30,40,50] for Car196. According to the results of Table 1, we select the H A with 30 superclasses for CIFAR100, and 15 for Car196.…”

“…However, these relations may be inconsistent with the appearances, which weakens the performance of classification tasks. Therefore, a lot of work builds visual information tree structures [13,3,19,20,21,16,11,29]. Some build the tree structures based on the confusion matrix [13,3,19,20,21], which is constructed by the results of a classifier.…”

“…Therefore, a lot of work builds visual information tree structures [13,3,19,20,21,16,11,29]. Some build the tree structures based on the confusion matrix [13,3,19,20,21], which is constructed by the results of a classifier. Others construct the label structure based on the affinity matrix [16,11,29] calculted by the similarity of any two categories.…”