Research On The Deep Network Model Of Small Target Detection And Recognition For Complex Scene

With the vigorous development of deep learning,target detection,as one of the most basic and challenging tasks in the field of computer vision,has achieved remarkable achievements.Current algorithms are mostly aimed at conventional targets with a certain size.However,the images captured in real scenes have small target sizes and unobvious features.Therefore,the detection performance of small targets is far from satisfactory.Small target detection,as a technology widely used in industrial production and life and military remote sensing scenes,has received extensive attention in recent years,and various methods have been proposed one after another,but the effects of small target detection are relatively poor.In this article,focusing on the problem of small target detection with few data samples and difficulty in extracting features,the cascade network model is adopted to realize the small target detection task.The main research contents are:(1)The current small target detection task is analyzed and its technical difficulties are explained.And the challenges faced in practical applications,and analyzed and introduced the solution and improvement of the small target detection algorithm and the related algorithm structure of the target detector;(2)Aiming at the problem of small targets in the image with small pixels and few data samples,this article An expansion method for small target dataset based on generative confrontation network is proposed.The experimental results show that the generative adversarial network has obvious sample expansion effects on helmet and small face data sets;(3)In the small target detection task,a cascade consisting of two parts of the detection module and the deep residual classification module is used Network model.The algorithm idea of this model first extracts the suspected small target area,and then performs specific classification and recognition.On the industrial helmet data set produced,the experimental results show that the cascade network model is two percentage points higher than the general target detection algorithm;(4)Looking forward to the future development trend of small target detection,and proposing a feasibility study algorithm,hoping to provide reference for the research work in this field.