Asphalt cracks on the highway surface are one of the most common pavement damage. If these cracks are not detected and taken care of on time, they grow and reach dimensions that will increase traffic safety and density. Therefore, it is very important to carry out the detection of asphalt cracks quickly. Traditional manual crack detection is extremely time consuming, very costly and requires a lot of effort. Therefore, the researchers concentrated their focus on the detection of automatic asphalt cracks. However, although automatic crack detection has been extensively investigated in recent years, it is still a difficult task due to the fact that cracks have different densities and the complexity of the pavement environment. In this study, a convolutional neural network-based method is proposed to overcome this difficulty. The proposed method was developed based on the convolution and inverted residual block structures used by MobileNetv2, which is known for its success and lightweight in classification and segmentation. As a result of the experimental tests, it is seen that the performance of the proposed method is higher than the other methods in the literature. This means that automatic asphalt crack detection is more successful.
Son yıllarda Türkiye’de zengin mineral, diyet lif ve vitamin içeren asma yapraklarının üretimi ve tüketimi yoğun olarak gerçekleşmektedir. Bununla birlikte hazır gıda sektöründe asma yapraklarından yapılan dolma yemeğine talep, farklı ülkelere ihracat olanaklarını da arttırmaktadır. Bunun gibi ticari tarım faaliyetlerinde sürdürülebilir bir pazarlama için kalite standartlarının oluşturulması önemlidir. Araştırmacılar, akıllı tarım uygulamalarında derin öğrenme ile birlikte olumlu ilerlemeler kaydetmiştir. Bu çalışmada, tüketim için kullanılacak asma yapraklarının türünün tanınması için yeni bir yöntem önerilmektedir. Önerilen yöntemde Ak, Ala Idris, Büzgülü, Dimnit ve Nazli olmak üzere 5 farklı asma yaprak türünden 500 görüntü içeren bir veri seti kullanılmıştır. Bu görüntülerden veri arttırma teknikleri ile 3500 adet görüntü elde edilmiştir. Ayrıca elde edilen görüntülere ESRGAN modeli uygulanarak daha ayrıntılı dokulardan oluşan bir veri kümesi elde edilmiştir. Bu görüntülerden öznitelik çıkarımı yapmak için VGG 19 derin öğrenme modeli kullanılmıştır. Oluşturulan iki ayrı veri setinden elde edilen öznitelikler birleştirilmiştir. Bu şekilde hibrit bir öznitelik çıkarıcı model oluşturulmuştur. PCA algoritması kullanılarak en iyi 175 adet öznitelik alt kümesi seçilmiştir. Son olarak elde edilen özniteliklerin sınıflandırılması için Destek Vektör Makinesi (DVM) kullanılarak %96,14 oranında doğruluk hesaplanmıştır.
In recent years, the usage areas of convolutional neural networks (CNN) have increased remarkably. It is widely used on many platforms, from workstations to embedded devices. However, each CNN model uses a different amount of memory, processor, storage and has different object recognition accuracy rates. CNNs to be used in embedded systems have some difficulties such as being less costly, consuming less resources and achieving higher accuracy. One of the CNN models that best overcomes these difficulties is the HBONet model. However, this model does not perform well enough in embedded systems. In this study, it is aimed to increase the performance of the HBONet model for embedded systems. For this purpose, the A-HBONet model, which is based on the HBONet model, is proposed. As a result of the experiments performed, the accuracy of the proposed model was increased by 3% compared to the HBONet model, while the memory and storage unit usage was reduced by approximately 80%. These results show that the proposed model works more effectively and efficiently in embedded devices.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.