Girish Mondal | Bangabandhu Sheikh Mujibur Rahman Science & Technology University, Bangladesh (original) (raw)

Uploads

Thesis Chapters by Girish Mondal

Research paper thumbnail of Fish classification using deep learning depending on shape and texture

Bangladesh, known for its abundance of rivers and ponds, has a rich cultural heritage centered ar... more Bangladesh, known for its abundance of rivers and ponds, has a rich cultural heritage centered around its love for fish, making it a renowned fish-loving nation, acquiring the name "Machh-e Bhat-e Bangali" (which means, Bengali by fish and rice). So, efficient monitoring of fish populations is essential for maximizing cultivation and promoting economic growth in our country, given the significant role that fishes play in the country's economy. Furthermore, Bangladesh's general economy might benefit from a strong fish monitoring system. So, in order to comprehend the strategies chosen, we evaluate several approaches and develop a system that can classify fishes. This study included two datasets. Dataset 1 which is our original dataset included 5 different native fish species and Dataset 2 which is collected included 8 different indigenous fish species from Bangladesh. The method uses segmentation, feature extraction, and ensembles as preprocessing steps to create the final result. In the preprocessing layer, U2-net is utilized to produce two different types of features: shaped pictures and colored images without backgrounds. The features were obtained using a feature descriptor that utilized transfer learning. Dataset 1 of 2678 fishes with 5 different classes, experimental results show accuracy of 99.72% for the first ensemble and 99.66% for the second ensemble layer. On the other hand, Dataset 2 of 2545 fishes with 8 different classes, experimental results show accuracy of 95.55% for the first ensemble and 97.341% for the second ensemble layer. The expected results across several levels were compared using a variety of performance measures.

Research paper thumbnail of Fish classification using deep learning depending on shape and texture

Bangladesh, known for its abundance of rivers and ponds, has a rich cultural heritage centered ar... more Bangladesh, known for its abundance of rivers and ponds, has a rich cultural heritage centered around its love for fish, making it a renowned fish-loving nation, acquiring the name "Machh-e Bhat-e Bangali" (which means, Bengali by fish and rice). So, efficient monitoring of fish populations is essential for maximizing cultivation and promoting economic growth in our country, given the significant role that fishes play in the country's economy. Furthermore, Bangladesh's general economy might benefit from a strong fish monitoring system. So, in order to comprehend the strategies chosen, we evaluate several approaches and develop a system that can classify fishes. This study included two datasets. Dataset 1 which is our original dataset included 5 different native fish species and Dataset 2 which is collected included 8 different indigenous fish species from Bangladesh. The method uses segmentation, feature extraction, and ensembles as preprocessing steps to create the final result. In the preprocessing layer, U2-net is utilized to produce two different types of features: shaped pictures and colored images without backgrounds. The features were obtained using a feature descriptor that utilized transfer learning. Dataset 1 of 2678 fishes with 5 different classes, experimental results show accuracy of 99.72% for the first ensemble and 99.66% for the second ensemble layer. On the other hand, Dataset 2 of 2545 fishes with 8 different classes, experimental results show accuracy of 95.55% for the first ensemble and 97.341% for the second ensemble layer. The expected results across several levels were compared using a variety of performance measures.