Research On Heterogeneous Embedded Fine-grained Image Automatic Classification System

With the development of deep learning,the performance of fine-grained image classification algorithm has been greatly improved.The fine-grained image classification algorithm is different from the traditional image recognition algorithm,it mainly judges the categories of different subclasses under the same category.Even experts in the field find it difficult to distinguish all the species because the differences between the targets are small and the parts of them are uncertain.For example,some birds differ only in the color of their wings.However,fine-grained image classification has a very wide range of application scenarios and research needs in many fields,but there are few hardware devices dedicated to fine-grained image classification in the market,so this paper will design a fine-grained image classification system to fill this need.It is difficult for common hardware equipment to meet the requirements of finegrained image classification algorithm.Heterogeneous embedded devices have the advantages of small size core structure,small size,low power consumption,high efficiency,customized software and hardware,etc.It is a very good choice to use heterogeneous embedded devices for fine-grained image classification.Therefore,this paper selects Rockchip's RK3399 heterogeneous embedded platform and implements fine-grained image classification system on this platform.In this paper,the whole fine-grained image classification system is divided into hardware part and software part.In the hardware part,the selection and construction of hardware platform including embedded processor,operating system and image acquisition module are completed.The software has completed the design and implementation of various modules including image preprocessing,fine-grained image classification and human-computer interaction.In the image acquisition module,the camera module OV13850 is selected as the image acquisition device.Then,the light supplement module is designed and implemented,and PWM is used to control the input of the light supplement module,so that users can adjust the brightness of the light supplement module according to the surrounding environment,which increases the practicability of the device.In the image preprocessing module,the common image noise reduction algorithm,image fog removal algorithm and image adaptive noise reduction algorithm based on morphology are designed and implemented.The image obtained from the image acquisition module is taken as the input of the image preprocessing part and the input image is processed by noise reduction in parallel in the three algorithms.Then the image sharpness after noise reduction is calculated respectively and the image with the highest sharpness after noise reduction is taken as the output of image preprocessing module.This paper focuses on the fine-grained image classification module.First of all,it deeply researches the Object-Part Attention Driven Discriminative Localization algorithm based on weak supervision,and introduces its basic principle,overall process,advantages and disadvantages in detail.One of them is the algorithm used to generate candidate regions in the original paper is modified to generate candidate regions more accurately and faster with less cost,and the target objects are guaranteed in each candidate region.Another one is designed a convolutional neural network model and implemented with the deep learning tool,which replaces the network model used in the original algorithm for target object positioning and feature extraction.Then the improved algorithm was trained and tested on the fine-grained image data set Caltech-UCSD Birs-200-2011 and Cars-196.Experimental results show that compared with the original algorithm,the improved algorithm in this paper improves the classification accuracy by 0.35%,reduces the training time by 7%~10%,and reduces the test time by 1~3S on the Caltech-UCSD Birds-200-2011 dataset.On the dataset Cars-196,the classification accuracy was improved by 0.37%,the training time was reduced by 8%~13%,and the test time was reduced by 2~4s.The human-computer interaction part is mainly based on QT design and implementation of GUI operation interface.Finally,the paper integrates the hardware part and the software part of the system,and connects the image acquisition module,image preprocessing module,fine-grained image classification module and human-computer interaction module in series on the hardware platform RK3399 to realize the fine-grained image automatic classification system based on heterogeneous embedded system.