Multilabel Image Classification using the CNN and DC-CNN Model on Pascal VOC 2012 Dataset

The recent widespread usage of image resources has led to the availability of a vast amount of data in for of images. Classifying or labelling images in an automated way is an ongoing research problem and categorizing data with several labels is still a challenging ask. Due to the intricacy of the data, it could be hard to infer information about classes that don’t mutually exclude one another. Many studies have concentrated on developing deep learning models for image classification, but the results do suffer due to poor low-frequency labelling accuracy and lack of fair distribution of the semantic images with annotations. This research study develops a dual-channel convolutional neural network (DC-CNN) model to increase the efficiency of automatic labelling. Two distinct CNN channels are included in the model of convolutional neural networks (CNN). To enhance the ratio of low-frequency samples, the first channel is used for training on low-frequency sampling, while the second channel is utilized for training on all training sets. While performing labeling, the results from the two channels are unifies to create a labelling decision. The suggested approach was validated using the Pascal VOC 2012 standard dataset. On just the Pascal VOC 2012 data sets, the proposed DC-CNN approach obtains an average maximum accuracy (AP) of above 95% for train data class. Additionally, for the train data class, the suggested CNN approach achieves 100% accuracy. Additionally, in terms of performance, our suggested model exceeds the conventional approach.