By examing the origin and development of Philosophy and Science,this paper proposes the Parallel Philosophy as a new way to view Artificial Intelligence,Intelligent Technology and related Smart Societies,based on Karl Popper′s Three Worlds Framework.This enables us to consider Being,Becoming,and Believing with Descriptive,Predictive,and Prescriptive knowledge,respectively.The mechanism and approach of integrating DAO in Blockchain and Monads in Mathematical Categories for implementing Intelligent Systems in line of the Parallel Philosophy are also investigated.

Homogeneous person re-identification(ReID) based on RGB images is well researched.With the development of public security and pedestrian retrieval in complex real-world situations,the field of pedestrian matching based on multi-modal heterogeneous data sources,called cross-modality heterogeneous ReID,is explored.In this paper,heterogeneous ReID and the differences from homogeneous ReID are firstly summarized.Then,heterogeneous ReID is described under six types of scenarios,including text-to-image,images-to-video,cross-resolution,infrared-visible,RGB-depth and sketch-photo.The relevant popular datasets,representative algorithms and excellent performance of these algorithms on datasets are summarized and reorganized.Finally,the current research progress and the future research trends of cross-modality heterogeneous person re-identification are discussed.

To improve the recognition accuracy of zero-shot recognition methods based on generative adversarial network(GAN),a zero-shot text recognition method combining transfer guidance and bidirectional cycle structure GAN is proposed.Bidirectional cycle structure GAN is constructed to improve the generation ability of the model,thus the generated pseudo features are closer to the real features of the input.The concept of transfer guided learning is introduced,and the model is trained by the transfer text instead of the seen text to improve the recognition accuracy of the unseen text.By adding an effective regularization term,the generator generates diverse results during the training process,and thus the stability of the generated model is improved.The experiment shows that the proposed method improves the accuracy of zero-shot recognition task with high generalization performance and it can be easily extended to other applications.

To improve the accuracy of footprint pressure image retrieval,a footprint pressure image retrieval algorithm based on multi-scale self-attention convolution is proposed.Firstly,preprocessing operations,such as angle correction,alignment and erasure,are carried out to reduce the influence of image angle on feature extraction.Secondly,the discriminative features are extracted adaptively by the multi-scale self-attention convolution module composed of the hole convolution with multiple parallel branches and the self-attention module.Finally,the common incomplete scoring matrix is obtained by the incomplete scoring module composed of global feature branches and incomplete score mask branches.The discriminative features are weighted and combined via the common incomplete scoring matrix to improve the attention of the network to the common visible area of incomplete footprints.The experimental results show that the proposed algorithm produces higher first hit accuracy and average retrieval accuracy on the constructed FootPrintImage dataset compared with some existing image retrieval methods.

In the object detection of thermal images,the image degradation phenomena,like the simple texture and the blurred object boundary,result in difficulties in localizing objects and matching the objects with the predefined anchor boxes.Therefore,an object detection algorithm for degraded thermal image based on feature alignment and assisted excitation of key points is proposed.Firstly,the visible image branch is introduced,and the similarity between the thermal domain and the visible domain is improved by calculating the feature difference of specified layers in two branches.Then,feature map concatenation and detection scale are modified to enrich the details of objects in the high-level network layers.Finally,an anchor-free detector with assisted excitation of key points is deployed,and thus the model localizes objects better and learn the instances poorly covered by the predefined anchor boxes.Comparative experiments on two datasets show that the proposed algorithm localizes thermal objects accurately and improves the accuracy of object detection in degraded thermal image effectively.

How to preserve the original features and reduce the distribution differences of multi-modal data more efficiently during the autoencoder learning process is an important research topic.A cross-modal retrieval approach via dual adversarial autoencoders(DAA) is proposed.The global adversarial network is employed to improve the data reconstruction process of the autoencoders.The min-max is implemented to make it difficult to distinguish the original features and reconstructed features.Consequently,the original features are preserved better.The hidden layer adversarial network generates modality-invariant representations and makes the inter-modal data indistinguishable from each other to reduce the distribution differences of multi-modal data effectively.Experimental results on Wikipedia and NUS-WIDE-10k datasets show the effectiveness of DAA.

Aiming at fabric defect detection with complex patterns,a fabric defect detection method based on distortion correction and visual salient features is proposed.Firstly,the image period is calculated to obtain the best block template,and then the image distortion is corrected according to the template.Secondly,the image is decomposed into texture layer and cartoon layer,and only the cartoon layer with the main features of the image is retained.Then,the improved context-aware saliency algorithm is applied to obtain the saliency feature of the image cartoon layer,so that the defects with high saliency features are separated from the background with low saliency features.Finally,the K-means clustering algorithm is utilized to highlight defects and complete defect detection.Experiments show that the proposed method achieves a high average recall rate for star,box and dot pattern fabrics,and the average recall precision effect of the proposed method is superior to that of the existing methods.

In defect location and classification of complex colored fabric,it is difficult to locate and classify defects in the cloth with complex and changeable background information.To solve this problem,a defect detection algorithm of complex pattern fabric based on cascaded convolution neural network is proposed.Firstly,the backbone feature extraction network based on two-way residual is applied to extract and fuse features from defect map and template map.Then,a density clustering frame producer is designed to guide the design of pre inspection frame for regional candidate networks in the framework.Finally,the cascaded regression method is utilized to locate and classify the defects accurately.The cloth image data collected from industrial field is adopted for training and prediction.The final results show that the proposed algorithm achieves high accuracy and recall rate.