Kyukwang Kim’s research while affiliated with Korea Advanced Institute of Science and Technology and other places

Mosquito control is important as mosquitoes are extremely harmful pests that spread various infectious diseases. In this research, we present the preliminary results of an automated system that detects the presence of mosquitoes via image processing using multiple deep learning networks. The Fully Convolutional Network (FCN) and neural network-based regression demonstrated an accuracy of 84%. Meanwhile, the single image classifier demonstrated an accuracy of only 52%. The overall processing time also decreased from 4.64 to 2.47 s compared to the conventional classifying network. After detection, a larvicide made from toxic protein crystals of the Bacillus thuringiensis serotype israelensis bacteria was injected into static water to stop the proliferation of mosquitoes. This system demonstrates a higher efficiency than hunting adult mosquitos while avoiding damage to other insects.

In image processing and robotic applications, two-dimensional (2D) black and white patterned planar markers are widely used. However, these markers are not detectable in low visibility environment and they are not changeable. This research proposes an active and adaptive marker node, which displays 2D marker patterns using light emitting diode (LED) arrays for easier recognition in the foggy or turbid underwater environments. Because each node is made to blink at a different frequency, active LED marker nodes were distinguishable from each other from a long distance without increasing the size of the marker. We expect that the proposed system can be used in various harsh conditions where the conventional marker systems are not applicable because of low visibility issues. The proposed system is still compatible with the conventional marker as the displayed patterns are identical.

Image-based sensing of jellyfish is important as they can cause great damage to the fisheries and seaside facilities and need to be properly controlled. In this research, we present a deep-learning-based technique to generate a synthetic image of the jellyfish easily with autoencoder-combined generative adversarial networks. The proposed system can easily generate simple images with smaller number of datasets compared to other generative networks. The generated output showed high similarity with the real image dataset. The application using a fully convolutional network and regression network to estimate the size of the jellyfish swarm was also demonstrated, and showed high accuracy during the estimation test.

Here, MineLoC is described as a pipeline developed to generate 3D printable models of master templates for Lab-on-a-Chip (LoC) by using a popular multi-player sandbox game “Minecraft”. The user can draw a simple diagram describing the channels and chambers of the Lab-on-a-Chip devices with pre-registered color codes which indicate the height of the generated structure. MineLoC converts the diagram into large chunks of blocks (equal sized cube units composing every object in the game) in the game world. The user and co-workers can simultaneously access the game and edit, modify, or review, which is a feature not generally supported by conventional design software. Once the review is complete, the resultant structure can be exported into a stereolithography (STL) file which can be used in additive manufacturing. Then, the Lab-on-a-Chip device can be fabricated by the standard protocol to produce a Lab-on-a-Chip. The simple polydimethylsiloxane (PDMS) device for the bacterial growth measurement used in the previous research was copied by the proposed method. The error calculation by a 3D model comparison showed an accuracy of 86%. It is anticipated that this work will facilitate more use of 3D printer-based Lab-on-a-Chip fabrication, which greatly lowers the entry barrier in the field of Lab-on-a-Chip research.

Microfluidic devices are an emerging platform for a variety of experiments involving bacterial cell culture, and has advantages including cost and convenience. One inevitable step during bacterial cell culture is the measurement of cell concentration in the channel. The optical density measurement technique is generally used for bacterial growth estimation, but it is not applicable to microfluidic devices due to the small sample volumes in microfluidics. Alternately, cell counting or colony-forming unit methods may be applied, but these do not work in situ; nor do these methods show measurement results immediately. To this end, we present a new vision-based method to estimate the growth level of the bacteria in microfluidic channels. We use Fast Fourier transform (FFT) to detect the frequency level change of the microscopic image, focusing on the fact that the microscopic image becomes rough as the number of cells in the field of view increases, adding high frequencies to the spectrum of the image. Two types of microfluidic devices are used to culture bacteria in liquid and agar gel medium, and time-lapsed images are captured. The images obtained are analyzed using FFT, resulting in an increase in high-frequency noise proportional to the time passed. Furthermore, we apply the developed method in the microfluidic antibiotics susceptibility test by recognizing the regional concentration change of the bacteria that are cultured in the antibiotics gradient. Finally, a deep learning-based data regression is performed on the data obtained by the proposed vision-based method for robust reporting of data.

Recently, owing to changes in weather conditions, cyanobacterial blooms, also known as harmful algal blooms (HABs) have caused serious damage to the ecosystems of rivers and lakes by producing cyanotoxins. In this paper, for the removal of HABs, an algal bloom removal robotic system (ARROS) is proposed. ARROS has been designed with a catamaran-type unmanned surface vehicle (USV) and an algae-removal device attached below. In addition, electrical control systems and a guidance, navigation, and control (GNC) system are implemented on the ARROS to remove the algal bloom autonomously. Moreover, to increase the efficiency of the work, an unmanned aerial vehicle (UAV) is utilized and the system detects algal blooms with an image-based detection algorithm, which is known as a local binary pattern (LBP). The overall mission begins with a command from a server when the UAV detects an algal bloom, and the USV follows the given path autonomously generated by a coverage path planning algorithm. Subsequently, with an electrocoagulation and floatation (ECF) reactor under the USV, HABs are removed. The performance of the algal bloom detection and HABs removal is verified through outdoor field tests in Daecheong Dam, South Korea.

Simple methods using the striped pattern paper marker and FFT (fast Fourier transformation) have been proposed as alternatives to measuring the optical density for determining the level of bacterial growth. The marker-based method can be easily automated, but due to image-processing-base of the method, the presence of light or the color of the culture broth can disturb the detection process. This paper proposes a modified version of marker-FFT-based growth detection that uses a light emitting diode (LED) array as a marker. Since the marker itself can emit the light, the measurements can be performed even when there is no light source or the bacteria are cultured in a large volume of darkly colored broth. In addition, an LED marker can function as a region of interest (ROI) indicator in the image. We expect that the proposed LED-based marker system will allow more robust growth detection compared to conventional methods.