Improving the performance of deep learning-based classification when a sample has various appearances

Jin, Guanghao; Jiao, Yuming; Wang, Jianming; Ma, Ming; Song, Qingzeng

doi:10.1080/0952813x.2022.2092558

Cited by 1 publication

(1 citation statement)

References 54 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The results obtained were not satisfactory for both image modalities. A second architecture has been proposed in order to perform results (Jin et al, 2022), especially when using IR images. This alternative is based on the fine-tuning of a pre-trained network.…”

Section: Introductionmentioning

confidence: 99%

A Novel Multispectral Maritime Target classification based on ThermalGAN (RGB-to-Thermal Image Translation)

Bouchenafa

Nemra

Amamra

et al. 2023

Journal of Experimental & Theoretical Artificial Intelligen

View full text Add to dashboard Cite

Convolutional Neural Networks (CNN) for ship classification in multi-spectral images (RGB, IR, etc.) is proposed in this paper. Recent developments in deep learning have significantly advanced the field of ship recognition. However, since maritime light intensity is frequently disturbed, multispectral imaging is considered a more robust substitute for RGB imaging. The proposed architectures were fine-tuned after being trained from scratch on the publicly available dataset VAIS (RGB-IR pairs). Unfortunately, the classification results plateaued at 59.74% accuracy, which is unsatisfactory for most real-life applications. Such an accuracy wall was due to the small number of training images. In order to overcome the scarcity of IR ship images, we proposed a novel image data augmentation strategy that translates RGB images to IR. A Pix2Pix model and a Generative Adversarial Network (GAN) network were modified to carry out the generation process as an RGB to IR translator. The KAIST general-purpose RGB-IR image pairs dataset was used to train our RGB-to-IR image translator, whereas the VAIS dataset was held aside for validation purposes. Our proposed network improved the accuracy of the native network by 8% ( from 59.74% to 67.74%), which is fairly satisfactory in the field of ship recognition.

show abstract

Section: Introductionmentioning

confidence: 99%