Detection System for Construction Image Classification Based on Deep Learning Models

Nowadays, image recognition and classification methods are widely used in various industries such as automatic driving, face recognitions. However, the classification of constructions is rarely studied in the past. In recent years, Image recognition and classification are mainly based on Convolutional neural networks (CNNs) which is one of the representative algorithms of deep learning. CNNs can greatly reduce the error rate of image classification. Therefore, we proposed a self-defined CNN to do image recognition and classification for some images of buildings, roads, bridges and so on. We collected pictures from Google image manually as our dataset. For our model, we first constructed the convolutional layer with filter size of 32 kernels, and then created a pooling layer to extract features. We repeated this process two more time and gradually adding the number of kernels to reduce sampling and the value of training. In the end, we constructed a fully-connected layer. Lastly, our method can achieve 0.2034 in loss, 0.9065 in accuracy, 0.8675 in validation loss and 0.7059 in validation accuracy, which is a relatively high level and acceptable in many real applications.

I. Introduction

With the development of artificial intelligence, the Image recognition technology has improved a lot, and it is widely used in our daily life, e.g. face recognition [1]-[2], speech recognition [3]-[4], image classification [5] and segmentation [6] as well as pose recognition [7]. Especially, it is also used to recognize the types of constructions including buildings, bridges, roads. In fact, rapid and accurate classification for constructions is needed. The objective of construction recognition based on deep learning models is that: to use data sets of construction images to train the model, so that we can realize that when we input a image of a construction, the computer can give the type of the construction accurately. Once we master the precise recognition of these constructions, we can make great progress in map drawing, digital city modeling, GPS and so on. Specifically, we can use satellites to obtain a large number of images of constructions on the earth. With the recognition of these images, we can get different types of constructions distributed in different geographical locations, so that we can realize map drawing, positioning or other operations. Thus, the recognition of constructions by image recognition is the basis of all these applications. Therefore, how to create an accurate, efficient and convenient recognition algorithm is an important research topic.