Materials and methods
1.1 Lighting effects on EEG and visual mechanism
Based on convolutional neural network models, an EEG learning network (EEGNet) was established. Because EEGNet is suited to processing EEG signals, it can be applied to EEG recognition. However, a bottleneck is that the EEG channel selection of a specific brain-computer interface (BCI) affects EEGNet recognition accuracy. This work develops an integrated EEGNet model for human visual mechanism light-intensity perception. First, a dedicated BCI was constructed using a designed multiplexer, EEG acquisition circuit, amplifier, and filter. Second, the constructed BCI was used to investigate the effect of ambient illumination on EEG. Finally, an integrated EEGNet was used to build a visual mechanism model. Experiments show that, compared with multi-channel EEGNet and single-channel EEGNet, the integrated EEGNet improves light-intensity recognition accuracy by 8.4% and 3.9%, respectively. The integrated EEGNet can effectively sense and classify ambient illumination intensity.
1.2 Light-intensity sensing principle and physical implementation
Humans perceive and recognize light intensity via visual organs. Figure 2 shows the principle and physical implementation of a light-intensity classifier. Implementing light-intensity perception requires hardware and software modules. The hardware includes a dedicated BCI, multiplexing circuitry, amplifiers, and filters. Software includes ambient illumination description, dedicated BCI control methods, EEG signal processing and feature extraction, and illumination recognition modeling. The light-intensity classifier module comprises model construction, training, testing, and real-time classification using the integrated EEGNet model. The physical implementation includes an electrode cap with a 16-channel EEG acquisition circuit, an 8-channel EEG acquisition circuit, EEG acquisition circuitry mounted on a 3D-printed helmet, and a helmet-mounted light-intensity classifier, as shown in Figure 3.
Figure 1. Visual mechanism and modeling
Figure 2. Light-intensity classifier principle and physical implementation
1.3 Dedicated brain-computer interface
For the dedicated BCI core function, a single-channel BCI was designed using the EEG sensing chip TGAM. The hardware logical structure is shown in Figure 4. Based on the TGAM chip, HC05 Bluetooth was used for EEG signal transmission. A controller, ARM NanoPi-Duo2, controls these circuits. The physical structure for multi-channel EEG is shown in Figure 5. A multiplexer selects one channel from multiple raw EEG channels. The selected channel is determined by the ARM controller NanoPi-Duo2. Compared with traditional BCIs, the specially constructed BCI can collect EEG from multiple electrodes via a single TGAM chip, producing more complete EEG images.
Figure 4. Dedicated BCI hardware logical structure
Figure 5. Multi-channel EEG sensing
For the dedicated BCI, the power circuit, EEG sensing circuit, and Bluetooth circuit were redesigned, as shown in Figure 6. All hardware chips use the power shown in (a). The EEG sensing chip TGAM and LED circuit are shown in (b). Communication between TGAM and the ARM controller uses the Bluetooth circuit shown in (c).
Figure 6. Circuit design. a. Power chip. b. EEG sensing chip. c. Bluetooth chip circuit for the dedicated BCI.
1.4 Construction of EEG amplifiers
Because EEG signals are weak and noisy, they need approximately 100x amplification. Figure 7 shows commonly used operational amplifiers and a two-stage op amp configuration for EEG amplification. A commonly used op amp was selected. Parameters such as gain, common-mode rejection ratio, and input impedance directly affect amplifier performance. The op amp was designed and parameter impacts were analyzed.
Figure 7. EEG amplifier. a. Common operational amplifiers. b. Two-stage EEG amplifier
1.5 Notch filter construction
After amplification, mains interference remains significant and negatively affects EEG analysis. Therefore, a notch filter is required. This filter passes most frequency information while attenuating a narrow rhythmic component. Figure 8 shows an active notch filter composed of a twin-T network. An op amp and potentiometer were added to the active notch filter. The op amp U2 provides feedback, and the potentiometer Rw is used for adjustment.
Figure 8. Active notch filter for EEG
Figure 9. Environmental light-induced EEG rhythm components μβ, μα, μθ, μδ and raw μο
1.6 EEG feature extraction
To model light intensity based on human visual perception, EEG signals are used as model inputs while illumination parameters are model outputs. Rhythmic components and approximate entropy were extracted from EEG as features. Figure 9 shows the raw EEG μο and rhythm components μβ, μα, μθ, and μδ. Wavelet transform was used to obtain the rhythmic energy components Eβ, Eα, Eθ, and Eδ, whose energy distributions are shown in Figure 10.
Figure 10. Energy feature distribution of EEG rhythm components
1.7 Integrated model for light-intensity sensing
Convolutional neural networks (CNN) are an efficient recognition model that performs well on large-scale image processing and primarily consist of convolutional and pooling layers in a feedforward structure. Because EEGNet adapts to EEG processing, it can be applied to EEG recognition. However, EEG selection from a specific BCI affects EEGNet accuracy. Therefore, an integrated EEGNet model for human visual mechanism light-intensity perception was developed, as shown in Figure 11. Multi-channel CNN and single-channel EEGNet extract raw EEG features in parallel. The model input and output must be determined. For EEG feature extraction, multi-channel CNN input is the EEG rhythmic components, and EEGNet input is the EEG connectivity matrix. The light-intensity classifier output corresponds to commonly used illumination categories: dim, mild, and bright. The multi-channel CNN structure includes multi-channel input, convolution, pooling, and fully connected layers.
Figure 11. Integrated EEGNet model for light-intensity sensing
The part (1) in Figure 11 is the multi-channel CNN model, which takes rhythmic components and EEG energy as input features. Part (2) is a single-channel EEGNet model with EEG connectivity matrix input. Part (3) is a radial basis function (RBF) network. Combining parts (1) and (2) in parallel covers both time-frequency and spatial features, yielding richer EEG information than common time-frequency features alone. The integrated EEGNet projects light-intensity features from low to high dimensionality, and the RBF network converts high-dimensional nonlinearly separable data into a low-dimensional linearly separable form, improving recognition accuracy.
This work used the designed BCI light-sensing system to explore the visual mechanism model. The multiplexing circuit, EEG acquisition circuit, amplifier, filter, and controller for the multi-channel BCI were designed and implemented. Ambient illuminance ranges were defined as 0-60 Lx, 61-120 Lx, and 121-350 Lx. EEG rhythm energy and entropy were extracted as features. Experiments show that the proposed integrated EEGNet improves light-intensity recognition accuracy by 8.4% and 3.9% compared with the single-channel EEGNet and multi-channel EEGNet models, respectively. Future work will further extract EEG features from connectivity matrices and optimize the BCI structure by reducing the number of lobes monitored.
