Research On Low Illumination Image Enhancement And Recognition Method Based On Capsule Network

Computer vision has already occupied a pivotal position in the field of artificial intelligence,which is leading the world's advanced technology.Image recognition has always been one of the hotly discussed and constantly conquered directions.At present,researchers are focusing on conventional image recognition and target detection.Excellent research results have been obtained in such tasks.However,in our daily work and life scenes,there are still many image-related tasks in extreme environments that are easily overlooked,such as image processing tasks in rainy,foggy,low-illuminance,and low-resolution conditions.For low-illumination scenes,due to the exposure of the shooting equipment and insufficient light in the real scene,the brightness of the obtained image is usually low,and the obtained image will lose a lot of effective information,and the visual effect is also relatively Poor,making subsequent image processing such as image recognition,detection,segmentation and other tasks ineffective.At present,most of the work related to low-illuminance images is to preprocess the low-illuminance image to enhance the brightness first,and then perform other subsequent tasks.However,the two independent tasks are too much in today's fast-paced and high-efficiency engineering requirements.It is cumbersome.Based on the particularity of low-illuminance images,this article has developed a method for low-illuminance image enhancement and a method for low-illuminance image recognition.The main contents are as follows:(1)Aiming at low-illuminance images with low brightness,uneven illumination,and inconspicuous image features,a low-illuminance image enhancement network based on mutual information variational capsule autoencoder is proposed.First,in the encoder,the mean value and variance are generated by obtaining the features of the capsule structure,and the latent variable z is obtained by using heavy parameters.In order to enrich the details of the generated enhanced image,input a latent variable c and latent variable z with a prior distribution for characteristics fusion.Then three decoding blocks composed of up-sampling and convolutional layers are used in the decoder to generate an enhanced image.In order to make the latent variable c have the greatest dependence on the output generated graph,mutual information maximization is introduced,and a Q module composed of four layers of convolution is added to fit the posterior probability of latent variable c during the maximization process.(2)Aiming at the problem of poor image recognition effect caused by the lack of low-illuminance image information,a low-illuminance image recognition algorithm based on attention mechanism and capsule network is proposed.By using global and local attention to acquire features for fusion,more identifiable information is obtained from low-illuminance images,and the weights obtained are weighted in the features to be identified,and combined with the capsule network structure.The experimental results show that,compared with other conventional image recognition algorithms,this algorithm directly solves the problem of low-light image recognition and obtains better experimental results.