Search

Purpose. Investigation of the possibilities of increasing the accuracy and variability of recognition of different groups of objects with similar redundant information by a neural network, after preprocessing and grouping of training images for further use on microcontrollers. The effect of removing redundant information in the training data on the practical values of recognition accuracy of different categories of objects based on the MobileNet V2 architecture is verified. Methodology. Creating different groups of training images using the Edge Impulse software platform. Image processing by the method of removing unnecessary objects in the graphic editor Adobe Photoshop. Testing the recognition accuracy of raw, processed and mixed groups of training images. Findings. Several groups of training images have been created on the basis of the Edge Impulse software platform. In part of the images, all unnecessary elements were removed using the Adobe Photoshop graphic editor. It was established that in the presence of similar redundant information on the training data, in different categories of objects, unrecognized and falsely recognized test images appear as a result of recognition. It was experimentally found that the method of removing redundant information from training images gives a clearer distribution of features, and the combination of raw and processed training data gives an average increase in recognition accuracy of more than 10% for each category, as well as a significant reduction of unrecognized and incorrectly recognized test images, while the same amount of training data. Originality. A combined method of training data processing and grouping was obtained, which increases the accuracy of object recognition without increasing the amount of training data. The effect of similar redundant information in different categories of objects on recognition accuracy was investigated. Practical value. The obtained results make it possible to increase the accuracy of recognition of different groups of objects with similar redundant information by one neural network without increasing the number of training images. [ABSTRACT FROM AUTHOR]

The development of information technologies, including artificial intelligence, is leading to their widespread use in many areas of our lives. With the rapid development of artificial intelligence, the number of cases of its application to generate digital images, audio, video, etc. is growing. Counterfeiting digital audio files is dangerous from the point of view of being used by criminals to exert information and psychological influence and manipulate society and individuals. There are modern methods for detecting audio fakes made by artificial intelligence. They are highly accurate, but not without drawbacks. The main drawback is the complex architecture and high resource intensity. The aim of this paper is to develop a neural network that would allow detecting audio fakes with satisfactory accuracy and whose training would not require significant computing resources. A method was chosen for modification, namely, the method of detecting audio fakes using a convolutional neural network. The method for detecting audio fakes was modified by building a model with a new architecture with fewer layers, which significantly reduced the time and computational resources required compared to analogues. The experimental results obtained using the modified method show satisfactory efficiency and accuracy of the system. Errors of the first kind amounted to 24%, errors of the second kind - 9%. The developed modification is implemented in a software application with a convenient and simple interface. [ABSTRACT FROM AUTHOR]

A rapid and accurate detection of a maize plant has been widely determined the food safety and quality assurance in modern agriculture. However, there is a great challenge on the complex environment of canopy leaves at the jointing stage. It is also necessary to select a feasible network model suitable for the precise identification and positioning of maize plants in the field. In this study, a novel device was developed for canopy recognition during the maize jointing stage using deep learning. A total of 3000 images of maize in the seedling and jointing stages were also taken to train the model of deep learning. Furthermore, a labeling strategy was proposed for the sample set, where the whole maize object was replaced with the center of the maize plant. Some data enhancement techniques were utilized to increase the robustness and generalization ability of the identification model for maize plants, including the passive brightness, chroma, rotation, mirroring, and blurring. After that, three datasets of maize plants were achieved, including 10 500 training sets, 3 000 cross-validation sets, and 300 test sets. In the model training, the SSDLite-MobileDet network model was first trained on the Google Colab cloud platform, and then compared with the SSDLite-MobileNetV3 model. Finally, an optimal SSDLite-MobileDet model was achieved with a speed of 110 frames/s and an accuracy of 92.4%. More importantly, some strategies were proposed to improve the performance of the model. 1) The specific procedure of extraction was visualized for the maize images during model training using a Convolutional Neural Network (CNN). It was found that the target information was significantly enriched in the output feature mapping, as the number of CNN layers deepened gradually. 2) The heart of the plant was labelled to avoid the serious leaf crossing at the jointing stage of maize. Thus, the novel model presented an average recognition accuracy of 92.4% under severe leaf crossing and different lighting conditions. 3) A rapid detection device was built for the maize crops with Raspberry Pi 4B+Coral USB, including Raspberry Pi 4B control, accelerator fast calculation, camera image acquisition, power, and image display module. As such, a rapid and accurate platform was obtained to collect, process, and display the images of the maize canopy in the field. 4) The model that trained on the GPU was quantized to transplant the trained model to the device. The data type of FP32 was converted to INT8, thereby ensuring that the quantized model occupied less memory. 5) A software was designed to run the SSDLite-MobileDet lightweight model on Raspberry Pi 4B, further to realize the acquisition, recognition, and display of maize images. The specific operation included system initialization, model loading, images acquisition, model processing, maize canopy recognition, positioning, and display. Finally, a field experiment was carried out to evaluate the rapid detection device for the maize crops, where the frame rate of the detected video was more than 89 frame/s, and the application recognition accuracy rate reached 91%. The findings can also offer strong support to the high-precision diagnosis and refined operation of maize in the field. [ABSTRACT FROM AUTHOR]