October 22, 2020
Our latest paper was featured in the international conference “ICPR 2020”
Adacotech Co., Ltd.(Head office: Chiyoda-ku, Tokyo, CEO: Ryota Kawamura, hereinafter referred to as “Adacotech”) is pleased to announce that a paper submitted by Adacotech has been selected for the International Conference on Pattern Recognition (ICPR 2020), which is scheduled to be held online from January 10 (Sunday) to 15 (Friday), 2021.
The content of the paper proposes a method for optimizing image preprocessing, which is essential for image recognition using HLAC(* 1) and deep learning.
This method allows us to automatically determine the best preprocessing method.
As a result, the workload of determining the processing method, which used to be carried by humans, is reduced, and optimal results can be achieved without depending on the skills of the engineer performing the processing.In addition, it is worth mentioning that it enables complex pretreatment that cannot be designed by hand.
Adacotech will continue to research and develop technologies related to AI-based image analysis and apply new technologies to its own solutions.We will promote the automation of inspection and testing by further promoting the implementation of our solutions in manufacturing sites. We also aim to develop businesses that support the evolution and innovation of the manufacturing industry.
* 1HLAC feature extraction method:A general-purpose and fast feature extraction method with excellent recognition accuracy used for image analysis and recognition.In contrast to the Deep Learning technology, which requires complex processing when calculating the shape and size of the inspection target, it can be calculated by simply performing a sum-of-product operation on the values representing the color shade and brightness of each pixel, which can be instantly calculated on a commercial PC.It is also capable of recognizing the same object even if its position has changed, and when there are two objects, the sum of the features of each object becomes the overall feature.In other words, there is no need to find the boundaries of regions that represent the same thing in an image (segmentation free), and the features can be recognized individually even when multiple anomalies occur in an image.It is a feature extraction method with desirable properties for image recognition.
AdaFilter: Adaptive Filter Design with Local Image Basis Decomposition for Optimizing Image Recognition Preprocessing
In recent years, the use of image processing to obtain information about objects has become very popular.For image recognition, methods using manual feature design, statistical analysis, and deep learning are used.In these methods, in order to achieve sufficient recognition performance, etc., it is important to make adjustments such as removing noise in the image before feature extraction and preprocessing by spatial filtering to emphasize components useful for recognition.Therefore, we investigated methods to optimize the preprocessing of the images.
In image processing, the method of finding an appropriate combination of typical image filtering processes by discrete optimization is already known.However, when the number of combination patterns is huge, manual design is difficult and automatic optimization is also difficult from the viewpoint of computational complexity.On the other hand, this method is based on the new idea of continuous optimization of the local image basis to determine the appropriate preprocessing filtering automatically and appropriately.By applying this method, the optimal preprocessing method is automatically selected, thus reducing the workload of determining the processing method.By applying this method, the optimal preprocessing method is automatically selected, thus reducing the workload of determining the processing method.
■「International Conference on Pattern Recognition」（ICPR2020）https://www.icpr2020