Abstarct: Capsule endoscopy is a non-invasive method for diagnosing diseases, which can assist gastroenterologists to investigate the digestive tract. This new technology has many advantages over conventional endoscopy. However, it is a tedious task for physicians to visually investigate the whole video frxames obtained in each endoscopy. Hence, a system is required to automatically detect probable frxames to any abnormalities for further investigation by the physician. Abnormality detection is challenging work. Different abnormalities such as bleeding, Angiodysplasia, polyp, and ulcer may exist in capsule endoscopy images. Each of these abnormalities can be detected by different features, including color, texture, or shape. Bleeding and angiodysplasia are usually distinguished only by the color property, but the ulcer regions are not only identified by color features and their texture differs from the normal regions. Three types of features, including color, texture, and shape, can be used to diagnose polyps. So detecting different abnormalities requires extracting the different types of features. Unfortunately, WCE does not provide image qualities equivalent to traditional methods. WCE also suffers from various impairments such as blurring, noise, and low illumination in different areas of the digestive tract. So, the details of the vessels and tissues are not clear in some frxames. In this research, a method has been proposed to investigate capsule endoscopy images for abnormalities including bleeding, ulcers, polyps, angiodysplasia, and lymphoid hyperplasia. The proposed method is a combination of deep and traditional learning techniques. Deep features are extracted from a pre-trained deep network. To extract hand-crafted features, at first, the image lighting is enhanced. Then, the potential suspected regions of abnormality are identified and from the obtained region, some features are extracted. Finally, by using extracted deep and hand-crafted features frxames are classified into mentioned lesions and normal classes. The results of this study on different datasets show that the proposed system is able to detect multiple lesions from WCE frxames with high accuracy and it has a better performance compared with other existing methods.